Deep neural architecture empowered by explainable artificial intelligence for accurate and early diagnosis of gynaecological cancer using medical images Authors: Sabah Abdullah Al-Somali, Zenah Mahmoud AlKubaisy, Muhyaddin Rawa, Khalid H. Allehaibi, Rania M. Alhazmi, Mahmoud Ragab Published: 2026 ABSTRACT Early analysis is a necessity for the more effective treatment of cancers. In gynaecological cancers, like endometrial, ovarian, and cervical cancers, the current efforts are aimed at discovering new analytical biomarkers to help decrease the global health burden related to these cancers. In cervical cancer, efficient screening is highly suggested for the prevention of invasive cancer occurrence and death. However, the interpretation of medical images for gynaecological cancer remains prone to human error. Artificial intelligence-based solutions offer numerous medical image challenges to help with the clinical decision support process. This paper presents a Deep Neural Architecture Empowered for Accurate Diagnosis of Gynaecological Cancer (DNAE-ADGC) model using medical imaging. The primary purpose of the paper is to empower gynaecological cancer diagnostics by developing an accurate, efficient, and intelligent detection framework using advanced techniques. Initially, the image pre-processing stage employs a two-level approach named adaptive filter that contains Median-Modified Wiener Filter (MMWF) and Cross Guided Bilateral Filter (CGBF). Followed by, the MobileNetV3Large model was deployed for feature extraction process. Besides, the DNAE-ADGC algorithm is applies the graph convolutional network and gated recurrent unit (GCN-GRU) network for detecting and classifying gynaecological cancer. At last, the explainable artificial intelligence (XAI) technique applies Grad-CAM to develop the transparency, interpretability, and reliability of AI models. The comparative analysis of the DNAE-ADGC method demonstrated an improved accuracy value of 97.92% with other methodologies under the Malhari dataset. Introduction As per recent statistics, cancer is the primary cause of death before the age of 70 years in several nations. Gynecologic cancers result from the uncontrolled propagation of abnormal cells in the female generative system, such as the uterus, cervix, ovaries, vagina, vulva, and fallopian tubes1. Cancer occurrence and related death rates differ widely between countries depending on the stage of economic progress and related cultural and behavioural factors. Among the different kinds of gynecologic cancers, breast cancer records the highest occurrence and death rates, followed by cervical cancer (CC)2. CC continues to be one of the most prevalent gynaecologic cancers and holds the second rank globally in terms of frequency and death rates. Globally, CC is the primary malignant tumour in women. There is a decline in the occurrence and death from CC in nations with an established screening system. CC Screening mainly relies on exfoliative cytology. At the same time, the Pap smear test can identify pre-invasive CC, thus significantly lowering the occurrence of invasive disease3. Although it is a reliable screening method, it cannot eliminate CC as the test lacks strong sensitivity. Currently, advanced techniques are introduced to enhance detection. Timely and accurate identification, along with suitable treatment methods, are crucial for a better prognosis and quality of life4. Yet, the evaluation of analytic images and pathology reports is intricate and subject to error. In several medical fields, the expert understanding and hands-on abilities that healthcare workers depend on involve recognising subtle patterns and abnormalities in medical images, such as radiological scans or histopathological slides5. Earlier research has indicated that many women seeking cancer screening are not aware of the most appropriate screening technique based on their age group. Research has also shown that even with insufficient awareness, women neither seek expert advice nor schedule gynaecological checkups. Most of these women receive a diagnosis of gynecologic cancers at a later stage6. Hence, informing women about gynecologic cancer signs may support early recognition and identification, which will make the treatment given more successful and possibly life-saving. Manual visual assessment involves a process that might depend on personal bias and irregularity in diagnosis, causing missed identifications and delayed care7. To address these issues, the utilisation of advanced computational techniques, DL, is designed to enhance the reliability and accuracy of gynaecological cancer recognition, becoming increasingly common and interesting8. A subset of artificial intelligence (AI) is DL, which has worked very effectively in several medical imaging tasks9. DL approaches in the area of gynaecological oncology can assist physicians in diagnosing and classifying cases with higher consistency and reliability, automatically extracting significant visual features from medical images of cases and merging numerous data types to support a diagnosis. Gynaecological cancers are diagnosed through DL, which can ideally increase the early and precise detection10. Therefore, they can support informed, personalised treatment methods and also improve patient outcomes. This paper presents a Deep Neural Architecture Empowered for Accurate Diagnosis of Gynaecological Cancer (DNAE-ADGC) model using medical imaging. The key contributions are listed below. A novel end-to-end framework is presented by integrating an advanced image pre-processing using MMWF and CGBF, efficient feature extraction with MobileNetV3Large, and accurate classification through a GCN integrated with GRU, along with Grad-CAM-based XAI to provide transparency, interpretability, and enhanced reliability in gynaecological cancer diagnosis.Apply a hybrid model integrating MobileNetV3 for feature extraction with GCN-GRU, introduced capturing spatial and relational dependencies. The transparency and clinical trust were improved by a Grad-CAM-based XAI. The proposed model attains greater performance compared to existing models, its effectiveness for reliable diagnosis for cancer diseases.The XAI technique using Grad-CAM is applied for generating visual explanations of the predictions. The transparency is improved, and critical regions in gynaecological images are detected. Furthermore, interpretability for clinicians is assisted by this process, and trust is built in automated diagnostic decisions while facilitating validation and comprehension of the DL model. A novel end-to-end framework is presented by integrating an advanced image pre-processing using MMWF and CGBF, efficient feature extraction with MobileNetV3Large, and accurate classification through a GCN integrated with GRU, along with Grad-CAM-based XAI to provide transparency, interpretability, and enhanced reliability in gynaecological cancer diagnosis. A novel end-to-end framework is presented by integrating an advanced image pre-processing using MMWF and CGBF, efficient feature extraction with MobileNetV3Large, and accurate classification through a GCN integrated with GRU, along with Grad-CAM-based XAI to provide transparency, interpretability, and enhanced reliability in gynaecological cancer diagnosis. Apply a hybrid model integrating MobileNetV3 for feature extraction with GCN-GRU, introduced capturing spatial and relational dependencies. The transparency and clinical trust were improved by a Grad-CAM-based XAI. The proposed model attains greater performance compared to existing models, its effectiveness for reliable diagnosis for cancer diseases. Apply a hybrid model integrating MobileNetV3 for feature extraction with GCN-GRU, introduced capturing spatial and relational dependencies. The transparency and clinical trust were improved by a Grad-CAM-based XAI. The proposed model attains greater performance compared to existing models, its effectiveness for reliable diagnosis for cancer diseases. The XAI technique using Grad-CAM is applied for generating visual explanations of the predictions. The transparency is improved, and critical regions in gynaecological images are detected. Furthermore, interpretability for clinicians is assisted by this process, and trust is built in automated diagnostic decisions while facilitating validation and comprehension of the DL model. The XAI technique using Grad-CAM is applied for generating visual explanations of the predictions. The transparency is improved, and critical regions in gynaecological images are detected. Furthermore, interpretability for clinicians is assisted by this process, and trust is built in automated diagnostic decisions while facilitating validation and comprehension of the DL model. Related works Bilal et al.11 proposed a new ensemble method, which employs the Salp Swarm Algorithm (SSA) to improve CC detection. To enhance feature attention, combined the Convolutional Block Attention Mechanism (CBAM) into every one of these methods. Then, the predictions made from the model in the ensemble are joined using a weighted aggregation model, and then the ensemble classifier is optimised by SSA. Umirzakova et al.12 introduced a lightweight and effective segmentation approach specially developed for cervical cell examination. This approach leverages a MobileNetV2 framework to extract features, guaranteeing a minimum parameter count favourable for practical processing. Moreover, incorporate extreme point supervision to improve segmentation results utilising minimum boundary annotation, instead of full pixel-wise labels. The authors13 presented a CAD-based CC identification through the Gazelle Optimiser Algorithm alongside a DL (CACCD-GOADL) technique. The main motive of this technique is to analyse the CC image. For this purpose, the presented technique utilises an enhanced MobileNetv3 paradigm to capture intricate shapes in Pap smear images. Furthermore, it creates a novel GOA for the parameter tuning of the improved MobileNet-v3 paradigm. Pacal14 developed a high-accuracy framework using the Multi-Axis Vision Transformer (MaxViT) model. The Mobile Bottleneck Convolution (MBConv) blocks are replaced with ConvNeXtv2 blocks, and the Multi-Layer Perceptron (MLP) blocks are replaced with Gated Residual Network (GRN)-based MLPs to improve generalisation and reduce parameters. Berahmand et al.15 proposed the Nested Ensemble DL for Gynaecological Cancer Prediction (NEDL-GCP) model by integrating neural networks (NNs) namely Convolutional NNs (CNNs), Recurrent NNs (RNNs), and Support Vector Machines (SVMs) as base learners, with J48 and Stochastic Gradient Descent (SGD) to refine predictions and improve classification accuracy. Muksimova et al.16 proposed RL-CancerNet, utilizes EfficientNetV2 for extracting fine-grained local features, Vision Transformers (ViTs) for capturing global dependencies, and a Reinforcement Learning (RL) agent to address class imbalance. Also, a Supporter Module with Convolutional 3D (Conv3D) and Bi-directional Long Short-Term Memory (BiLSTM) layers integrated with attention mechanism improves contextual feature learning. The authors17 presented an Improved Bald Eagle Search Optimiser alongside a DL-driven CC Detection and Classification (IBESODL-CCDC) method. This method aims to automate the detection and classification of CC. In this method, a contrast enhancement procedure is carried out to improve the image quality. Furthermore, this method employs an adapted LeNet methodology for the feature extraction. To diagnose CC, this method utilises an attention-based long short-term memory (ALSTM) method. Sahoo et al.18 developed a model by using three pretrained models to predict CC early. To combine these model results, a new fuzzy rank-based ensemble is considered by dual nonlinear functions. In contrast to fusion approaches, this framework presents the latest predictions by taking into account the base classifier’s certainty in forecasts. To additionally improve the classification abilities of these methods, merge sophisticated augmentation methodologies. Mishra et al.19 created an automatic CC prediction model that relies on Pap smear images, which utilises a pooling model. In CC detection, image examination is crucial. With developing techniques, computer vision (CV) and CNNs are of prior significance due to their unbelievable ability to deal with object detection and image identification difficulties. Almujally et al.20 utilised Transfer Learning (TL) with the MobileNet CNN and Chi-square feature selection (FS) to detect the most impactful features. Kodipalli, Fernandes, and Dasar21 proposed a model by utilising CNNs incorporated with Transformers and an Ensemble MLP along with classical ML models for classification. Abdel-Salam, Askr, and Hassanien22 utilised AI and DL methods by integrating pre-trained models for feature extraction, Copula Entropy-based Golden Jackal Optimisation (CE-based GJO) for feature selection, Sea Horse Optimiser (SHO) for hyperparameter tuning, and Explainable AI (XAI) for classification. Hussain et al.23 proposed a method by utilising XAI and seven XAI methods, including Saliency, XRAI, Integrated Gradients, Smooth and Integrated Gradients, Grad- and Smoothgrad-CAM, applied to four CNN models, such as AlexNet, VGG16, MobileNet, and InceptionNet, to evaluate model predictions and align them with human expert annotations. Kodipalli et al.24 proposed a model by utilising XAI techniques. A three-stage ensemble method combined with a game-theory model based on SHapley Additive exPlanations (SHAP) is also used to analyse and visualise the most important features. AlMohimeed et al.25 utilised ViT for feature extraction, Particle Swarm Optimization (PSO) for FS and optimization, and SVM for precise classification. Additionally, XAI via Grad-CAM is employed to interpret model predictions and highlight essential image regions. Nazir et al.26 employed AI, ML, and DL approaches on holographic microscopic imaging data. It integrates pre-trained DL models for feature extraction, classical ML classifiers for classification, and XAI approaches such as Grad-CAM and LIME to interpret model decisions and enhance diagnostic transparency. The authors27 examined a model by implementing a Bilinear Convolutional NN (BCNN) integrated with a Random Projection (RP) paradigm. The approach also employs dyadic feature extraction and bilinear pooling to generate rich composite image representations. Saha et al.28 utilised the OvCan-FIND model, which employs CNNs for feature extraction and comparison with multiple advanced CNN architectures. Comparison analysis of existing Gynaecological Cancers Detection in Table 1. Table 1A comparative study of diverse sophisticated models.ReferencesYearObjectivesMethodsDatasetResultsBilal et al.112025To present an innovative ensemble model for combining three sophisticated pre-trained CNNs.SSA, CBAMMendeley LBC and BloodMNIS DatasetsAccuracy of 99.48% and 95.23%.Umirzakova et al.122025To propose an effective and lightweight segmentation method, particularly intended for cervical cell identification.MobileNetV2SipakMed DatasetDice Coefficient 0.87 and F1 Score 0.84.Nour et al.132024To introduce a novel model for CC identification.CACCD-GOADL, MobileNetv3, GOAHerlev DatasetAccuracy 99.38%, Precision 96.71%, Recall 97.45%.Pacal142024To develop a high-accuracy, lightweight DL model.MaxViT, MBConv → ConvNeXtv2 Blocks, MLP → GRN-based MLP, Lightweight Architecture AdaptationSIPaKMeD Pap Smear, Mendeley LBC DatasetAccuracy: 99.02% (SIPaKMeD), Accuracy: 99.48% (LBC)Berahmand et al.152025To predict gynaecological cancer risk from Pap smear images using a nested ensemble DL model.NEDL-GCP, CNN, RNN, SVM, J48 + SGDHerlev and SIPaKMeD Pap SmearAccuracy 99.1% (Herlev), Accuracy 98.5% (SIPaKMeD)Muksimova et al.162024To enhance cervical cancer diagnosis accuracy using a hybrid DL model.EfficientNetV2, ViT, RL, Conv3D, BiLSTMHerlev, SipaKMeDAccuracy of 99.7%Al Mazroa et al.172023To allow for identifying relevant frameworks and patterns inside cervical images.IBESODL-CCDC, LeNet, ALSTM–Accuracy 99.26% and Precision 97.51%.Sahoo et al.182023Tp present an innovative fuzzy rank-based ensemble method to utilise dual nonlinear functions of diverse concavities for managing the fuzzy ranks of classes.Fuzzy Rank-based EnsembleSIPaKMeD and Mendeley LBC DatasetsAccuracy of 97.18%, 99.22% and F1 score of 97.16%, 99.19%.Mishra et al.192023To project a hybrid pooling method.Pooling Model, CV, CNNHerlev and IARC DatasetsAccuracy of 95.45%.Almujally et al.202025For the timely and accurate detection of ovarian cancer.TL, MobileNet CNN and Chi-squareSoochow University Ovarian Cancer DatasetAccuracy: 98.49%, Precision: 99.13%, Recall: 99.27%, F1-Score: 99.20%Kodipalli, Fernandes, and Dasar212024To detect and classify ovarian cancer.CNN, Ensemble MLPOvarian CT Scan Images and Biomarker DataAccuracy: 98.96%, Precision: 97.44%, F1-Score: 98.7%, Dice Score: 0.98, Jaccard Score: 0.97Abdel-Salam, Askr, and Hassanien222025To develop an accurate and explainable AI-based model.CE-based GJO, SHO, XAISipakMed DatasetAccuracy: 0.9985, Precision: 0.9985, Specificity: 0.9996, F-Measure: 0.9985Hussain et al.232024To evaluate XAI techniques.AlexNet, VGG16, MobileNet, InceptionNet, Saliency, XRAI, Grad-CAM, Integrated, Smooth, and Fusion Gradients, Smoothgrad-CAMCervical Dysplasia Image DatasetSensitivityKodipalli et al.242025To develop an accurate and explainable AI-based system.XAI, SHAPMulti-Modal Real-Life Ovarian Cancer DataAccuracy of 98.66%AlMohimeed et al.252024To accurately predict cervical cancer.ViT, PSO, SVM, Grad-CAMSipakMed and Herlev Cervical Cell ImagesAccuracy: 99.11% (SipakMed), 97.78% (Herlev); F1-score: 99.11% (SipakMed), 97.81% (Herlev)Nazir et al.262025To develop an accurate and explainable AI-based framework.Pre-trained DL, SVM, RF, KNN, DT, AdaBoost, GBM, Grad-CAM, LIMEHolographic and Bright-Field Cervical Cancer ImagesAccuracy of 100%, Precision, Recall, F1-Score, and AUC-Score of 1.00, MSE and RMSE of 0.00Abd-Alhalem et al.272024To improve CC classification accuracy and efficiency.BCNN, RPCervical Cell Image DatasetAccuracy: 0.9983 (dual-label), 0.9530 (multiclass)Saha et al.282025To develop an accurate and robust model.CNN, OvCan-FINDOvarian Cancer Image DatasetAccuracy of 99.74% Table 1A comparative study of diverse sophisticated models.ReferencesYearObjectivesMethodsDatasetResultsBilal et al.112025To present an innovative ensemble model for combining three sophisticated pre-trained CNNs.SSA, CBAMMendeley LBC and BloodMNIS DatasetsAccuracy of 99.48% and 95.23%.Umirzakova et al.122025To propose an effective and lightweight segmentation method, particularly intended for cervical cell identification.MobileNetV2SipakMed DatasetDice Coefficient 0.87 and F1 Score 0.84.Nour et al.132024To introduce a novel model for CC identification.CACCD-GOADL, MobileNetv3, GOAHerlev DatasetAccuracy 99.38%, Precision 96.71%, Recall 97.45%.Pacal142024To develop a high-accuracy, lightweight DL model.MaxViT, MBConv → ConvNeXtv2 Blocks, MLP → GRN-based MLP, Lightweight Architecture AdaptationSIPaKMeD Pap Smear, Mendeley LBC DatasetAccuracy: 99.02% (SIPaKMeD), Accuracy: 99.48% (LBC)Berahmand et al.152025To predict gynaecological cancer risk from Pap smear images using a nested ensemble DL model.NEDL-GCP, CNN, RNN, SVM, J48 + SGDHerlev and SIPaKMeD Pap SmearAccuracy 99.1% (Herlev), Accuracy 98.5% (SIPaKMeD)Muksimova et al.162024To enhance cervical cancer diagnosis accuracy using a hybrid DL model.EfficientNetV2, ViT, RL, Conv3D, BiLSTMHerlev, SipaKMeDAccuracy of 99.7%Al Mazroa et al.172023To allow for identifying relevant frameworks and patterns inside cervical images.IBESODL-CCDC, LeNet, ALSTM–Accuracy 99.26% and Precision 97.51%.Sahoo et al.182023Tp present an innovative fuzzy rank-based ensemble method to utilise dual nonlinear functions of diverse concavities for managing the fuzzy ranks of classes.Fuzzy Rank-based EnsembleSIPaKMeD and Mendeley LBC DatasetsAccuracy of 97.18%, 99.22% and F1 score of 97.16%, 99.19%.Mishra et al.192023To project a hybrid pooling method.Pooling Model, CV, CNNHerlev and IARC DatasetsAccuracy of 95.45%.Almujally et al.202025For the timely and accurate detection of ovarian cancer.TL, MobileNet CNN and Chi-squareSoochow University Ovarian Cancer DatasetAccuracy: 98.49%, Precision: 99.13%, Recall: 99.27%, F1-Score: 99.20%Kodipalli, Fernandes, and Dasar212024To detect and classify ovarian cancer.CNN, Ensemble MLPOvarian CT Scan Images and Biomarker DataAccuracy: 98.96%, Precision: 97.44%, F1-Score: 98.7%, Dice Score: 0.98, Jaccard Score: 0.97Abdel-Salam, Askr, and Hassanien222025To develop an accurate and explainable AI-based model.CE-based GJO, SHO, XAISipakMed DatasetAccuracy: 0.9985, Precision: 0.9985, Specificity: 0.9996, F-Measure: 0.9985Hussain et al.232024To evaluate XAI techniques.AlexNet, VGG16, MobileNet, InceptionNet, Saliency, XRAI, Grad-CAM, Integrated, Smooth, and Fusion Gradients, Smoothgrad-CAMCervical Dysplasia Image DatasetSensitivityKodipalli et al.242025To develop an accurate and explainable AI-based system.XAI, SHAPMulti-Modal Real-Life Ovarian Cancer DataAccuracy of 98.66%AlMohimeed et al.252024To accurately predict cervical cancer.ViT, PSO, SVM, Grad-CAMSipakMed and Herlev Cervical Cell ImagesAccuracy: 99.11% (SipakMed), 97.78% (Herlev); F1-score: 99.11% (SipakMed), 97.81% (Herlev)Nazir et al.262025To develop an accurate and explainable AI-based framework.Pre-trained DL, SVM, RF, KNN, DT, AdaBoost, GBM, Grad-CAM, LIMEHolographic and Bright-Field Cervical Cancer ImagesAccuracy of 100%, Precision, Recall, F1-Score, and AUC-Score of 1.00, MSE and RMSE of 0.00Abd-Alhalem et al.272024To improve CC classification accuracy and efficiency.BCNN, RPCervical Cell Image DatasetAccuracy: 0.9983 (dual-label), 0.9530 (multiclass)Saha et al.282025To develop an accurate and robust model.CNN, OvCan-FINDOvarian Cancer Image DatasetAccuracy of 99.74% A comparative study of diverse sophisticated models. The existing DL Models for gynaecological cancer diagnosis mostly focus on accuracy but lack robustness for medical images in noisy handling. The data handled with many approaches relies only on CNN-based algorithms and spatial relationships that are complex to capture. Furthermore, most methods work with black-box techniques without providing proper interpretability, which limits clinical trust. The advanced hybrid models that were integrated were extracted for limited usage. Additionally, the image quality could not be enhanced by pre-processing techniques due to insufficient attention. Later, it is necessary to combined structure expands accuracy, expandability, and robustness for initial and dependable analysis of gynaecological cancers. Proposed materials This paper presents a DNAE-ADGC model using Medical Imaging. The primary purpose of the paper is to empower gynaecological cancer diagnostics by developing an accurate, efficient, and intelligent detection framework using advanced techniques. The DNAE-ADGC method involves image pre-processing phase, feature representation model, detection, and an XAI method. Figure 1 depicts the complete working flow process of the DNAE-ADGC method. Unlike traditional hybrid models that loosely integrate CNN, GCN, and RNN techniques, the presented architecture proposes a structured and task-specific incorporation: advanced preprocessing employing Median-Modified Wiener Filter and Cross-Guided Bilateral Filter to improve medical image quality, efficient feature extraction via MobileNetV3Large tailored for low-complexity yet high-representation learning, a tightly coupled GCN-GRU mechanism, which acquires both spatial relationships and sequential dependencies in feature maps, and the incorporation of XAI (Grad-CAM) to ensure clinical interpretability. This end-to-end design, integrates optimized preprocessing, lightweight feature extraction, graph-based relational learning, and explainability within a single unified pipeline, distinguishes the proposed model from previous works and establishes its novelty. Fig. 1Overall working flow of the DNAE-ADGC model. Overall working flow of the DNAE-ADGC model. Image pre-processing Initially, a two-level approach named an adaptive filter, which contains MMWF and CGBF is employed to enhance the image quality, contrast and by removing noise. The integration effectively mitigates noise while also conserving crucial edge and fine details in Pap smear images. MMWF effectively handles impulsive and Gaussian noise more robustly, compared to standard filters, and contrast and structural data are improved by CGBF by employing guidance from cross-channel data, resulting in more explicit images that will enhance extraction and subsequent accuracy. In the pre-processing stage, the method utilises a 2-level adaptive filtering tactic to increase image quality and permit for reliable analysis. The MMWF is initially applied to reduce Gaussian noise and impulse but maintaining edges, incorporating the advantages of the Wiener filter (WF) and the median filter (MF). Then, the CGBF was used to increase image contrast and maintain essential details by considering intensity differences and spatial closeness, as trained by a different image. These two methods guarantee that the images are well‐defined and clean, crucial to segment and extract. MMWF This model is applied to decrease the noise dispersion in degraded imaging. By utilising an MF for denoising the background area of a flawed image and suppress impulsive noise when preserving edges by replacing each pixel with the median of its neighborhood. Furthermore, this model utilising the WF to reduce noise while maintaining important structural details29. By replacing median values for the mask pixel values, the noise is reduced in the declining image. The median value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\stackrel{\sim}{\mu\:}\right)$$\end{document} substitutes for an average value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\mu\:\right)$$\end{document} in the WF computation. Therefore, the MMWF is stated in Eq. (1) as demonstrated:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b_{mmwf}}\left( {n,{\mathrm{~}}m} \right)=\tilde {\mu }+\frac{{{\sigma ^2} - {v^2}}}{{{\sigma ^2}}} \cdot \left( {a\left( {n,{\mathrm{~}}m} \right) - \tilde {\mu }} \right){\mathrm{~}}$$\end{document} Whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}^{2}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}^{2}$$\end{document} depicts the Gaussian- and noise variance set. In the mask, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\:$$\end{document}x m signifies the neighbourhood area size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document}, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a(n,\:m)$$\end{document} characterises every area pixel. The merit is that the image quality of the deteriorated image is heightened, as shown: The drop-off result maintains the edge signal, improved over the MF and WF techniques. CGBF CGBF is an edge-preserving smoothing technique, and it can remove artefacts without affecting the surface quality, preserving image quality and edge. The key characteristics of CGBF is maintains sharp edges while reducing noise and it uses both spatial weights and range weights from the guidance image. The CGBF is used in the recommended structure due to its merits. Similar to a bilateral filter (BF), the CBF preserves borders by considering geometric familiarity and grayscale similarity among neighboring pixels. Using one image for the filter and another for the kernel allows exploration of relationships among nearby pixels. The CGBF prevents gradient reversal and other elementary BF problems. Utilising the CBF approach, targeted at the input image I, the guided weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({w}_{g}\right)$$\end{document} is multiplied by the guided image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(F\right)$$\end{document}. Equation (2) describes the CGBF.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{P} = \frac{1}{W}\sum\limits_{{q \in \:S}} {G_{{\sigma_{s} }} } \left( {\left\| {P - q} \right\|} \right)G_{{\sigma_{r} }} w_{g} \left| {X\left( p \right) - X\left( q \right)} \right|I_{P}$$\end{document} Whereas W is well-defined utilising the succeeding Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W= \sum\limits_{q\in\:S}{G}_{{\sigma\:}_{s}}\left({\left\| {P - q} \right\|}\right){G}_{{\sigma\:}_{r}}{w}_{g}\left|X\left(p\right)-X\left(q\right)\right|$$\end{document} There are dual weighted parameters applied here: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{{\sigma\:}_{r}}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{{\sigma\:}_{s}}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{{\sigma\:}_{r}}$$\end{document} is calculated utilising the guided image. The spatial and range (colour/intensity) kernels are Gaussian-shaped and are described using Eqs. (4) and (5):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{{\sigma\:}_{s}}=\left({\left\| {P - q} \right\|}\right)={\mathrm{e}}^{-\frac{{{\left\| {P - q} \right\|}}^{2}}{2{\sigma\:}_{s}^{2}}}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G_{{\sigma _{r} }} = \left( {\left| {X\left( p \right) - X\left( q \right)} \right|} \right) = {\mathrm{e}}^{{ - \frac{{\left\| {X\left( p \right) - X\left( q \right)} \right\|^{2} }}{{2\sigma_{r}^{2} }}}}$$\end{document} The standard deviations (SD) are represented by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{s}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{r}$$\end{document}, whereas \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\left(q\right)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\left(p\right)$$\end{document} signify the guided image pixel colours of p and q, correspondingly. The spatial kernel weights are computed according to the Euclidean distance among q and p. S refers to the spatial domain. The CGBF is subject to the spatial and range kernels that are optimally chosen. Feature extraction using MobileNetV3Large model Followed by, MobileNetV3Large technique is employed for extraction30. This assists in attaining high accuracy with low computational cost, due to its squeeze-and-excitation blocks and hard-swish activations. Unlike conventional models such as Visual Geometry Group (VGG) or ResNet, MobileNetV3Large is effective for real-time or resource-constrained tasks. For GCN input, feature maps from a convolutional layer are extracted, and each spatial location is treated as a node. Spatial proximity or feature similarity is used for establishing connections (edges) between nodes, presenting an adjacency matrix. In contrast, the node-feature matrix captures the feature vectors, together serving as the input to the GCN. MobileNet-V3 Large is one of the most promising DL models intended for mobile and other edge gadgets. It depends on the brought-out principles of the preceding MobileNet framework and combines benefits like automated ML. The structural design is modified to accomplish greater execution on mobile hardware. This is formed by the depth-wise separable convolution in the network, which results in reducing the parameter counts, thus making the network more effective with fewer parameters. Furthermore, the model describes novel activation functions, such as the hard‐swish (h‐swish) and the hard‐sigmoid function, which are less computationally demanding compared with other tasks by enhancing efficacy. Self‐supervision and dynamic adjustment: The network employs a squeeze‐and‐excitation (SE) optimiser configuration, which allows it to modify channel‐wise feature responses usefully. This method will be able to pay attention to the most relevant features for enhancement. The depth-wise separable convolution is considered to be the factorisation of the basic convolution operation into the depth-wise and point-wise convolutions. This factorisation decreases the computation counts required. MobileNet-V3 Large integrates hard‐sigmoid and h‐swish activations to enhance efficacy. H‐swish presents non‐linearity with minimum computation cost, although hard‐sigmoid improves feature recalibration in the block of SE. The significance of MobileNet-V3 Large cannot be overstated. To employ prevailing DL techniques on edge devices, particularly in real‐time scenarios, namely medical imaging, having a method that preserves a balance among precision and computation efficacy is crucial. MobileNet-V3 Large, with its unique structural novelties and optimisers, excels as an ideal candidate for this application, promising strong performance without wide-ranging computation resources. The hard-sigmoid activation and h‐swish functions are particularly utilised in MobileNet-V3 Large, enhancing its efficacy for edge computing applications. The b‐swish function is a computationally efficient alternative to the standard swish activation, presents non‐linearity, while upholding a low computation footprint. It is employed in a depth-wise convolution layer, contributing to quicker convergence and enhanced precision without considerably increasing the inference period. Assume that X depicts an input tensor with dimensions H×W×C, where H, W, and C refer to height, width, and channels. The standard convolution operation was described below:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Y=X\mathrm{*}K$$\end{document} Now * represents a convolution operation, and K indicates a convolutional filter with dimensions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{H}\times\:{k}_{W}\times\:C\times\:M$$\end{document}, with M being the output channel count. In depth-wise separable convolution, this operation is factorised into:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{dw}=X\star\:{K}_{dw}$$\end{document} Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\star\:$$\end{document} depicts depth-wise convolution, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{dw}$$\end{document} being a filter of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{H}\times\:{k}_{W}\times\:C$$\end{document}. This is succeeded by point-wise convolution:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{pw}={Y}_{dw}\odot\:{K}_{pw}$$\end{document} Now \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\odot\:$$\end{document} represents point-wise convolution with a filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{pw}$$\end{document} of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:1\times\:C\times\:M.$$\end{document} This factorisation decreases the cost of computation from H \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} W \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{H}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}_{W}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} C \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\times\:$$\end{document} M to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H\times\:W\times\:({k}_{H}\times\:{k}_{W}\times\:C+C\times\:M)$$\end{document}. The h-swish function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h-swish\left(x\right)$$\end{document} and hard‐sigmoid function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h-sigmoid\left(x\right)\:$$\end{document}are specified:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h-swish\left(x\right)=x\cdot\:\frac{\mathrm{m}\mathrm{a}\mathrm{x}\left(0,min\left(x+\mathrm{3,6}\right)\right)}{6}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h-sigmoid\left(x\right)=\frac{\mathrm{m}\mathrm{a}\mathrm{x}\left(0,min\left(x+\mathrm{3,6}\right)\right)}{6}$$\end{document} These are first-order approximations of swish and sigmoid functions, respectively. The SE block is mathematically depicted as:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Z=\sigma\:\left({W}_{2}\cdot\:ReLU\left({W}_{1}\cdot\:S\left(X\right)\right)\right)$$\end{document} Now \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S\left(X\right)$$\end{document} represents global average pooling of X, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{2}$$\end{document} are learnable weighted matrices, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} indicates the activation function of sigmoid. The output Z is employed to scale the input tensor X channel-wise. Hybrid classification model Besides, the DNAE-ADGC model employs the GCN-GRU model to detect and classify gynaecological cancer31. Both spatial and temporal relationships in the data are used for this process. The spatial feature maps separated from the CNN are reformulated as graph-structured representations, whereas nodes respond to discriminative regions and edges acquire contextual relationships among them. The GCN efficiently models these non-Euclidean spatial dependencies, when the GRU acquires sequential and context-based features over graph nodes, allowing greater representation of complex tissue patterns. This integrated system allows both spatial correlation learning and contextual dependency modeling, which is not attainable by standalone techniques. In non-Euclidean structures of data, different traditional CNNs. GCN develops feature demonstration by combining neighbourhood data. In this model to learned complex data for global contextual information, which improved classification performance related to conventional models. GCN improves completing performance by preserving important information in the data. While GRU models sequential or contextual patterns efficiently. GRU are gating mechanisms for feature representations in capturing sequential dependencies more effectively. They are more efficient with minimal parameters to compare another recurrent methods, reduce the overfitting risk. The GRUs enable information flow and context retention to be better, which one donates to better classification performance. This integration outperforms standalone CNNs or RNNs by incorporating feature interactions and temporal dynamics, improving classification accuracy in complex medical datasets. The baseline method combines dual major elements: GCN and GRU, to control spatio-temporal relations. The GCN unit tackles spatial relations to combine features of neighbouring nodes inside a graph. Here, nodes are utilised for spatial relations. By employing a normalised adjacency matrix and learnable weights, the GCN layer effectually acquires spatial patterns inside the network. Every node has specified data and edges among nodes with a specified adjacency matrix A, both physical proximity and functional linkage. Adjacency and matrix have a set of features X for every node, equivalent to a set of features. The specific mechanism of the graph convolution procedure comprises neighbourhood feature aggregation that emphasises how spatial relations are encoded through the graph framework. Hence, the representation of an adjacency matrix A and the node feature X on a single layer of GCN creates a spatial representation.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H=\sigma \left( {A \cdot X \cdot W} \right)~$$\end{document} Here, X refers to the input feature matrix for nodes, A depicts the normalised adjacency matrix, H denotes the output feature matrix after convolution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} is the activation function, and W depicts the trainable weight matrix. To accomplish this, the GRU regulates temporal dependency to utilise update and reset gates for learning adaptively in time-sequential data. The reset gate controls previously existing hidden data, whereas the upgrade gate determines that new data will be integrated into the state. This structure allows the GRU to manage a meaningful flow of data from one time step to another in such a manner that it is capable to learn short- and long‐term temporal dependency effectively. The capability of GRU to forget unnecessary data from previous, while retaining essential data safeguards preceding and existing data. The efficacy of GRU in sequential data modelling with few parameters compared with regular LSTM. RNN is utilised to manage sequential data to learn temporal dependency through time. Internally, GRU utilises diverse gates for the regularisation of data flow that surpass the normal RNN. Their major internal gates, which are reset and update gates, allow it to select only the significant data from the previous or discard much of the unnecessary data, therefore improving essential patterns. GRU is simpler to design compared to LSTM because it does not have a cell state from the hidden layer (HL), which makes it computationally inexpensive while capable of accomplishing better performance for tasks that require long‐term dependency in sequence data to be modelled. Figure 2 represents the GRU structure. Fig. 2GRU structure. GRU structure. Update gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\boldsymbol{Z}}_{\boldsymbol{t}}\right)$$\end{document} This is the most significant unit in a GRU, which allows the replacement of data from a preceding time-step to the existing time-step in a sequence. It signifies how much of the preceding HL is required to perform the existing time-step and how much novel data is necessary to combine from the existing input. The upgrade gate acquires how much data from the preceding time-step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({h}_{t-1}\right)$$\end{document} must impact the existing state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(h\right)$$\end{document}, which is specified:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{t}=\sigma\:\left({W}_{z} \cdot\:\left[{x}_{t},\:{h}_{t-1}\right]+{b}_{z}\right)$$\end{document} Now, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}$$\end{document} signified input at the existing time-step, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{z}$$\end{document} depicts the weighted matrix for the upgrade gate, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{z}$$\end{document} is the bias term, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} refers to the sigmoid activation function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-1}$$\end{document} depicts HL from the preceding time-step. Reset gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\boldsymbol{r}}_{\boldsymbol{t}}\right)$$\end{document} Another significant unit of GRU is the reset gate that manages the amount of preceding HL to be ignored or forgotten while calculating the novel HL of a candidate. Although the upgrade gate dictates the amount of preceding HL to keep, the reset gate commands how much the preceding should enter the existing time step to generate a novel state of the candidate. It signifies how much preceding available data to throw away by computing the HL of the candidate is specified:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}=\sigma\:\left({W}_{r}\cdot\:\left[{x}_{t},\:{h}_{t-1}\right]+{b}_{r}\right)$$\end{document} Now, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{r}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{r}$$\end{document} refer to the weighted matrix and bias for the reset gate, respectively. Candidate hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\stackrel{\sim}{\boldsymbol{h}}}_{\boldsymbol{t}}\right)$$\end{document} The candidate HL of GRU is a novel HL at some time step before being incorporated with the present HL to present the final HL. It performs a pivotal part in novel data that the method is set to present based on input taken at this period and adapted preceding HL once influenced by the resetting gate:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{h} _{t} = {\mathrm{tan}}h\left( {W_{h} \cdot \:\left[ {x_{t} ,\:\left( {r_{t} \odot h_{{t - 1}} } \right)} \right] + b_{h} } \right)$$\end{document} Now, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{h}$$\end{document} signifies a biased term, ⊙ represents element-wise multiplication, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{h}$$\end{document} implies a weighted matrix for the HL of the candidate. Final hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\boldsymbol{h}}_{\boldsymbol{t}}\right)$$\end{document} The final HL of GRU is significant to acquire sequential data dynamics since it encapsulates data from preceding and existing inputs. This is organised by a balance between retention and upgrade, managed by the upgrade gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{t}$$\end{document} that dynamically modifies the number of contributions that come from the previous HL and assessed candidate HL. While the upgrade gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{t}$$\end{document} is adjacent to 1, this method keeps the preceding HL, which means the last data is highly relevant. The value of upgraded gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{t}$$\end{document} approaches 0; this method increasingly depends on the HL of the candidate to generate the final HL. The final HL integrates HL \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{h}}_{t}$$\end{document} and preceding HL \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({h}_{t-1}\right)$$\end{document} utilising the upgrade gate.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h_{t} = \left( {1 - z_{t} } \right) \odot h_{{t - 1}} + z_{t} \odot \underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle-}$}}{h} _{t}$$\end{document} This equation safeguards the method to retain valuable preceding data while upgrading with the more relevant existing data. XAI using the Grad-CAM technique At last, the XAI technique applies Grad-CAM to develop the transparency, interpretability, and reliability of AI models to make comprehension easy of humans. With the improved usage of DL, trustworthiness and transparency of the model are essential32. As a result, the XAI method Grad-CAM presents insights into how a trained CNN carries classification forecasts by recognising important graphic areas that impact the decision-making of the technique. Grad-CAM is an extensively applied interpretation method in DL, mainly suitable for visualising the internal decision process of CNN methods. Here, the training results of the model structure are examined utilising Grad-CAM-created heatmaps, which identified the critical areas in the input image that affected the classification decisions. This model was created by choosing the last network’s convolutional layer, as this layer maintains important spatial features while summarising higher-level forms necessary for classification tasks. After this, an automated difference was applied to calculate the output class score gradients regarding the chosen feature mapping layer. This gradient imitates how variation in the activation map influences the confidence of the method in its prediction. Then, a global average pooling process is used on these gradients, giving significance weights for all feature maps. These calculate the impact of feature mapping on the last output. To generate the heatmap, a weighted incorporation is computed, highlighting areas with a significant contribution to the result. The function of ReLU is used in this phase to eliminate negative values. The resulting heatmap is normalized to the 0–1 scale for consistency in intensity demonstration. The heatmap is colour-coded utilising OpenCV, where areas are underlined in yellow or red, signifying stronger activation. Then, this colored heatmap is applied to the new image to present a rich representation of regions mainly responsible for the decision of the method. This visualisation pipeline is used for each testing instance in the database, and the created heatmaps are restored for detailed examination. This model allowed a thorough analysis of the reasoning of the method, presenting insights into feature significance and improving the interpretability of the model. The significant weights asked for all feature maps, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}^{k}$$\end{document} as in Eq. (17).17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{k}=\frac{1}{Z}\sum\limits_{i}\sum\limits_{j}\frac{\partial\:{y}^{c}}{\partial\:{A}_{ij}^{k}}$$\end{document} Now, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}^{c}$$\end{document} signifies the score for the targeted class c, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{A}_{ij}^{k}$$\end{document} denote the spatial position activation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(i,\:j)$$\end{document} in feature mapping k, and Z characterises the complete feature mapping counts. The last feature mapping of Grad-CAM \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{Grad-CAM}^{c}$$\end{document} is gained over Eq. (18).18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{Grad-CAM}^{c}=ReLU\left(\sum\limits_{k}{\alpha\:}_{k}{A}^{k}\right)$$\end{document} The activation of the ReLU guarantees that only positively contributing areas are highlighted. Therefore, this model aids in recognising the particular portions of the image that most affected the prediction of the model, increasing transparency and confidence of the results. Result analysis The model was implemented in Python 3.11.0 and executed on a system with an Intel i5-8600 K processor, GeForce GTX 1050Ti (4GB), 16GB RAM, 250GB SSD, and 1 TB HDD. The training–validation protocol has been standardized with the following parameters: learning rate of 0.001, batch size of 8, dropout rate of 0.5, 50 training epochs, and ReLU activation function. The validation of the DNAE-ADGC model is tested under the Malhari dataset 33, which comprises 318 images under four classes, as presented below in Table 2. The dataset was divided using a 70:30 split, whereas 70% of the data was used for training and 30% for validation. This dataset is chosen for its labelled gynaecological cytology images, covering multiple cancerous and precancerous classes, making it appropriate for early diagnosis research. Also, data augmentation, class-weighted loss, and careful regularisation are utilised for ensuring robust model training and improved generalisation. Figure 3 represents the sample images. Figure 4 defines the sample images of HSIL, LSIL, NILM, and SSC. Table 2Dataset details.ClassImagesHSIL40LSIL80NILM158SSC40Total318 Table 2Dataset details.ClassImagesHSIL40LSIL80NILM158SSC40Total318 Dataset details. Fig. 3Sample images. Sample images. Fig. 4Sample images of HSIL, LSIL, NILM, and SSC. Sample images of HSIL, LSIL, NILM, and SSC. Fig. 5Sample of pre-processed images. Sample of pre-processed images. Figure 5 shows the sample outcome of the pre-processing step. Figure 6 shows the classifier outputs of the DNAE-ADGC approach on 80% and 20%. Figure 6a and b illustrate the confusion matrix. Figure 6c demonstrates the PR inspection, denoting maximal performance on each class. Ultimately, Fig. 6d exhibits the ROC examination. Fig. 680:20 of (a–b) confusion matrices and (c–d) curves of PR and ROC. 80:20 of (a–b) confusion matrices and (c–d) curves of PR and ROC. Table 3; Fig. 7 portray the cervical cancer recognition of the DNAE-ADGC method on 80% and 20%. Under 80% TRPHE, the DNAE-ADGC approach gets an average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 94.09%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 87.27%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 82.53%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}\:$$\end{document}of 84.51%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{AUC}_{Score}$$\end{document}of 88.77% and MCC of 80.14%. Likewise, at 20% TSPHE, the DNAE-ADGC model gets an average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 97.66%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 93.04%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 93.04%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}\:$$\end{document}of 93.04%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{AUC}_{Score}\:$$\end{document}of 95.72% and MCC of 91.44%. Table 3Cervical cancer recognition of DNAE-ADGC algorithm with 80:20.Class labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}\boldsymbol{u}\boldsymbol{r}\boldsymbol{a}\boldsymbol{c}\boldsymbol{y}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}\boldsymbol{c}\boldsymbol{i}\boldsymbol{s}\boldsymbol{i}\boldsymbol{o}\boldsymbol{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}\boldsymbol{l}\boldsymbol{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MCCTRPHE (80%) HSIL96.0688.4676.6782.1487.6680.20 LSIL97.2492.1996.7294.4097.0792.62 NILM89.7687.6893.0890.3089.6879.64 SSC93.3180.7763.6471.1980.6968.07 Average94.0987.2782.5384.5188.7780.14TSPHE (20%) HSIL96.8890.0090.0090.0094.0788.15 LSIL100.00100.00100.00100.00100.00100.00 NILM96.8896.4396.4396.4396.8393.65 SSC96.8885.7185.7185.7191.9883.96 Average97.6693.0493.0493.0495.7291.44 Table 3Cervical cancer recognition of DNAE-ADGC algorithm with 80:20.Class labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}\boldsymbol{u}\boldsymbol{r}\boldsymbol{a}\boldsymbol{c}\boldsymbol{y}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}\boldsymbol{c}\boldsymbol{i}\boldsymbol{s}\boldsymbol{i}\boldsymbol{o}\boldsymbol{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}\boldsymbol{l}\boldsymbol{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MCCTRPHE (80%) HSIL96.0688.4676.6782.1487.6680.20 LSIL97.2492.1996.7294.4097.0792.62 NILM89.7687.6893.0890.3089.6879.64 SSC93.3180.7763.6471.1980.6968.07 Average94.0987.2782.5384.5188.7780.14TSPHE (20%) HSIL96.8890.0090.0090.0094.0788.15 LSIL100.00100.00100.00100.00100.00100.00 NILM96.8896.4396.4396.4396.8393.65 SSC96.8885.7185.7185.7191.9883.96 Average97.6693.0493.0493.0495.7291.44 Cervical cancer recognition of DNAE-ADGC algorithm with 80:20. Fig. 7Average outcomes of the DNAE-ADGC system with 80:20. Average outcomes of the DNAE-ADGC system with 80:20. Figure 8 defines the training (TRAIN) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} and validation (VALID) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of a DNAE-ADGC methodology under 80% and 20% over 50 epochs. This emphasizes the robust ability in learning and retaining features. Fig. 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Acc{u}_{y}$$\end{document} curve of DNAE-ADGC system with 80% and 20%. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Acc{u}_{y}$$\end{document} curve of DNAE-ADGC system with 80% and 20%. Figure 9 demonstrates the TRAIN loss and VALID loss of the DNAE-ADGC model on 80:20 over 50 epochs. To begin with, the both loss is high but decrease steadily, depicting effective learning without overfitting. Fig. 9Loss curve of DNAE-ADGC model under 80:20. Loss curve of DNAE-ADGC model under 80:20. Figure 10 portrays the classifier outputs of the DNAE-ADGC algorithm on 70:30. Figure 9a and b display the confusion matrix. Figure 9c presents the PR, denoting maximal value on 4 classes. However, Fig. 9d elucidates the ROC, representing efficacious outputs with increased ROC values. Fig. 1070:30 of (a–b) confusion matrices and (c–d) curves of PR and ROC. 70:30 of (a–b) confusion matrices and (c–d) curves of PR and ROC. Table 4; Fig. 11 exemplify the cervical cancer detection of DNAE-ADGC system at 70:30. Under 70% TRPHE, the DNAE-ADGC approach reaches an average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 95.50%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 93.15%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 89.54%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}\:$$\end{document}of 91.08%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{AUC}_{Score}$$\end{document}of 92.85% and MCC 87.59%. Also, at 30% TSPHE, the DNAE-ADGC system attains an average \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 97.92%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 95.87%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 94.91%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}\:$$\end{document}of 95.20%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{AUC}_{Score}\:$$\end{document}of 96.64%, and MCC 93.78%. Table 4Cervical Cancer detection of DNAE-ADGC model under 70:30.Class Labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}\boldsymbol{u}\boldsymbol{r}\boldsymbol{a}\boldsymbol{c}\boldsymbol{y}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}\boldsymbol{c}\boldsymbol{i}\boldsymbol{s}\boldsymbol{i}\boldsymbol{o}\boldsymbol{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}\boldsymbol{l}\boldsymbol{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MCCTRPHE (70%) HSIL99.1096.5596.5596.5598.0296.03 LSIL93.2485.4587.0486.2491.1481.77 NILM91.8990.6093.8192.1791.8683.82 SSC97.75100.0080.7789.3690.3888.75 Average95.5093.1589.5491.0892.8587.59TSPHE (30%) HSIL98.9691.67100.0095.6599.4195.18 LSIL97.9296.1596.1596.1597.3694.73 NILM96.8895.6597.7896.7096.9393.75 SSC97.92100.0085.7192.3192.8691.47 Average97.9295.8794.9195.2096.6493.78 Table 4Cervical Cancer detection of DNAE-ADGC model under 70:30.Class Labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}\boldsymbol{u}\boldsymbol{r}\boldsymbol{a}\boldsymbol{c}\boldsymbol{y}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}\boldsymbol{c}\boldsymbol{i}\boldsymbol{s}\boldsymbol{i}\boldsymbol{o}\boldsymbol{n}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}\boldsymbol{l}\boldsymbol{l}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MCCTRPHE (70%) HSIL99.1096.5596.5596.5598.0296.03 LSIL93.2485.4587.0486.2491.1481.77 NILM91.8990.6093.8192.1791.8683.82 SSC97.75100.0080.7789.3690.3888.75 Average95.5093.1589.5491.0892.8587.59TSPHE (30%) HSIL98.9691.67100.0095.6599.4195.18 LSIL97.9296.1596.1596.1597.3694.73 NILM96.8895.6597.7896.7096.9393.75 SSC97.92100.0085.7192.3192.8691.47 Average97.9295.8794.9195.2096.6493.78 Cervical Cancer detection of DNAE-ADGC model under 70:30. Fig. 11Average outcomes of the DNAE-ADGC model with 70% and 30%. Average outcomes of the DNAE-ADGC model with 70% and 30%. Figure 12 depicts the TRAIN \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} and VALID \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of a DNAE-ADGC methodology at 70:30 over 50 epochs. This highlights superior results between TRAIN and VALID. Fig. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Acc{u}_{y}$$\end{document} curve of DNAE-ADGC system at 70% and 30%. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Acc{u}_{y}$$\end{document} curve of DNAE-ADGC system at 70% and 30%. Figure 13 defines the TRAIN loss and VALID loss of the DNAE-ADGC model at 70% and 30% over 50 epochs. Primarily, losses are high but steadily lessen, with closely aligned curves, illustrating effective learning and good generalization without overfitting. Fig. 13Loss curve of DNAE-ADGC model under 70:30. Loss curve of DNAE-ADGC model under 70:30. Table 5; Fig. 14 represent the comparative analysis of DNAE-ADGC methodology with current approaches under various metrics [35–37. The outputs underscored that the DNAE-ADGC system got the optimum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l},$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 97.92%, 95.87%, 94.91%, and 95.20%, respectively. While the other models, such as MaxViT, MLP, NEDL-GCP, ResNet50, XGB, SVM, Naïve Bayes (NB), KNN, Xception, and DeepLab v3, illustrated worst performance under diverse metrics. Table 5Comparison assessment of DNAE-ADGC model with existing methods.ModelAccuracyPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MaxViT94.1489.6492.5288.24MLP88.4588.1392.8493.26NEDL-GCP88.2290.2493.3193.94ResNet5094.4488.7294.2488.16XGB93.3889.0991.9987.56SVM87.8987.3393.2492.73NB87.5489.7093.6294.30KNN96.7993.6291.5089.81Xception95.9593.6792.2991.97DeepLab v396.0092.2787.4289.22DNAE-ADGC97.9295.8794.9195.20 Table 5Comparison assessment of DNAE-ADGC model with existing methods.ModelAccuracyPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MaxViT94.1489.6492.5288.24MLP88.4588.1392.8493.26NEDL-GCP88.2290.2493.3193.94ResNet5094.4488.7294.2488.16XGB93.3889.0991.9987.56SVM87.8987.3393.2492.73NB87.5489.7093.6294.30KNN96.7993.6291.5089.81Xception95.9593.6792.2991.97DeepLab v396.0092.2787.4289.22DNAE-ADGC97.9295.8794.9195.20 Comparison assessment of DNAE-ADGC model with existing methods. Fig. 14Comparison assessment of the DNAE-ADGC model with existing methods. Comparison assessment of the DNAE-ADGC model with existing methods. Table 6; Fig. 15 indicate the comparison assessment of the DNAE-ADGC system with other approaches16,34. The ResNet50 model attained an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 96.24%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 93.52%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 91.37%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 93.58%, DeiT achieved the best overall performance with strong \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document}. ViTs and MobileNetV2 performed slightly lower but remained consistent. EfficientNetV1 and ConvNext also depicted high effectiveness, with ConvNext close to DeiT in overall results. Finally, the DNAE-ADGC model outperformed all others with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 97.44%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 95.06%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 94.06%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 94.11%, emphasizing its superior capability in detecting subtle diagnostic features in cervical cytology images. Table 6Comparison outcome of the DNAE-ADGC approach with other methods.Methods \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} ResNet5096.2493.5291.3793.58DeiT97.0090.7392.1490.99ViTs94.4493.0292.0091.81EfficientNetV195.2191.8593.0193.45ConvNext96.5493.8890.6793.31MobileNetV295.2190.5190.3890.77DNAE-ADGC97.4495.0694.0694.11 Table 6Comparison outcome of the DNAE-ADGC approach with other methods.Methods \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} ResNet5096.2493.5291.3793.58DeiT97.0090.7392.1490.99ViTs94.4493.0292.0091.81EfficientNetV195.2191.8593.0193.45ConvNext96.5493.8890.6793.31MobileNetV295.2190.5190.3890.77DNAE-ADGC97.4495.0694.0694.11 Comparison outcome of the DNAE-ADGC approach with other methods. Fig. 15Comparison outcome of the DNAE-ADGC methodology with other models. Comparison outcome of the DNAE-ADGC methodology with other models. The execution time (EXT) of the DNAE-ADGC algorithm with existing approaches is proven in Table 7; Fig. 16. The DNAE-ADGC model presents a lower EXT of 3.52 min, while the MaxViT, MLP, NEDL-GCP, ResNet50, XGB, SVM, NB, KNN, Xception, and DeepLab v3 methodologies obtain a higher EXT of 9.22 min, 10.08 min, 8.56 min, 7.58 min, 6.73 min, 7.77 min, 6.54 min, 8.51 min, 6.87 min, and 8.50 min, respectively. Table 7EXT outcome of DNAE-ADGC model with recent methods.ModelEXT (min)MaxViT9.22MLP10.08NEDL-GCP8.56ResNet507.58XGB6.73SVM7.77NB6.54KNN8.51Xception6.87DeepLab v38.50DNAE-ADGC3.52 Table 7EXT outcome of DNAE-ADGC model with recent methods.ModelEXT (min)MaxViT9.22MLP10.08NEDL-GCP8.56ResNet507.58XGB6.73SVM7.77NB6.54KNN8.51Xception6.87DeepLab v38.50DNAE-ADGC3.52 EXT outcome of DNAE-ADGC model with recent methods. Fig. 16EXT outcome of DNAE-ADGC model with recent methods. EXT outcome of DNAE-ADGC model with recent methods. Table 8; Fig. 17 depict the error analysis of the DNAE-ADGC method with existing approaches. The MaxViT achieved an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 5.86%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 10.36%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 7.48%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 11.76%. The MLP and NEDL-GCP methods depicted low overall performance. Conventional models like ResNet50, XGB, SVM, and NB also performed poorly, with modest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document}. KNN, Xception, and DeepLab v3 had slightly better recall in some cases, but overall results remained low across all approaches. The DNAE-ADGC model achieved the minimal error rates with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 2.08%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 4.13%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 5.09%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 4.80%, highlighting its superior ability. Table 8Error analysis of the DNAE-ADGC approach with other systems.Model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MaxViT5.8610.367.4811.76MLP11.5511.877.166.74NEDL-GCP11.789.766.696.06ResNet505.5611.285.7611.84XGB6.6210.918.0112.44SVM12.1112.676.767.27NB12.4610.306.385.70KNN3.216.388.5010.19Xception4.056.337.718.03DeepLab v34.007.7312.5810.78DNAE-ADGC2.084.135.094.80 Table 8Error analysis of the DNAE-ADGC approach with other systems.Model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MaxViT5.8610.367.4811.76MLP11.5511.877.166.74NEDL-GCP11.789.766.696.06ResNet505.5611.285.7611.84XGB6.6210.918.0112.44SVM12.1112.676.767.27NB12.4610.306.385.70KNN3.216.388.5010.19Xception4.056.337.718.03DeepLab v34.007.7312.5810.78DNAE-ADGC2.084.135.094.80 Error analysis of the DNAE-ADGC approach with other systems. Fig. 17Error analysis of the DNAE-ADGC model with existing approaches. Error analysis of the DNAE-ADGC model with existing approaches. Table 9 specifies the efficiency evaluation of the DNAE-ADGC method37. The RetinaNet model required 24.88 Giga Floating Point Operations (FLOPs), 1952 MB of Graphics Processing Unit (GPU) memory, and an inference time of 6.74 s. YOLOv3-tiny used 27.65 G FLOPs, 2192 MB GPU, and 9.55 s for inference, while YOLOX-tiny required 24.41 G FLOPs, 2322 MB GPU, and 8.95 s. ESOD-YOLO attained the highest computational load with 29.62 G FLOPs, 2134 MB GPU, and 13.92 s inference time. ResNet50 attained 22.8 G FLOPs, 2406 MB GPU, and 4.79 s inference time, and the RF model required 27.78 G FLOPs, 2243 MB GPU, and 12.36 s. The DNAE-ADGC model illustrated superior efficiency with only 12.55 G FLOPs, 569 MB GPU memory usage, and an inference time of 3.98 s, emphasizing a significant reduction in computational cost while maintaining high performance. Table 9Computational analysis of the DNAE-ADGC methodology in terms of FLOPs, GPU, and inference time.MethodologyFLOPs (G)GPU (M)Inference time (sec)RetinaNet24.8819526.74YOLOv3-tiny27.6521929.55YOLOX-tiny24.4123228.95ESOD-YOLO29.62213413.92ResNet5022.824064.79RF27.78224312.36DNAE-ADGC12.555693.98 Table 9Computational analysis of the DNAE-ADGC methodology in terms of FLOPs, GPU, and inference time.MethodologyFLOPs (G)GPU (M)Inference time (sec)RetinaNet24.8819526.74YOLOv3-tiny27.6521929.55YOLOX-tiny24.4123228.95ESOD-YOLO29.62213413.92ResNet5022.824064.79RF27.78224312.36DNAE-ADGC12.555693.98 Computational analysis of the DNAE-ADGC methodology in terms of FLOPs, GPU, and inference time. Table 10 demonstrates the K-fold cross validation results obtained by the proposed model. The table values demonstrated that the proposed model attains effectual performance under each fold in terms of different measures. Table 10K-fold cross validation of the DNAE-ADGC algorithm.K-fold cross validationAccuracyPrecisionRecallF1-scoreCV 190.3987.5286.1787.43CV 285.7690.0785.5493.12CV 387.4389.1091.0585.70CV 488.1288.9085.3886.85CV 587.5893.8591.7688.49CV 692.0393.9887.4588.45CV 792.4689.2491.9892.01CV 888.6987.0487.8390.65CV 991.8288.8787.5885.88CV 1086.9485.9488.4687.69 Table 10K-fold cross validation of the DNAE-ADGC algorithm.K-fold cross validationAccuracyPrecisionRecallF1-scoreCV 190.3987.5286.1787.43CV 285.7690.0785.5493.12CV 387.4389.1091.0585.70CV 488.1288.9085.3886.85CV 587.5893.8591.7688.49CV 692.0393.9887.4588.45CV 792.4689.2491.9892.01CV 888.6987.0487.8390.65CV 991.8288.8787.5885.88CV 1086.9485.9488.4687.69 K-fold cross validation of the DNAE-ADGC algorithm. Table 11 illustrates the ablation study analysis of the DNAE-ADGC approach. The DNAE-ADGC method attains greatest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 97.92%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 95.87%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 94.91%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 95.20%, establishing the proficiency of Hybrid GCN-GRU+MobileNetV3Large with preprocessing techniques, while MobileNetV3Large+MMWF without CGBF showed an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 94.78%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 92.64%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 91.63%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 91.82%. Likewise, MobileNetV3Large+CGBF without MMWF improved metrics to an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 95.34%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 93.40%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 92.22%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 92.56%. Additionally, GCN only without GRU and feature extraction achieved an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 95.99%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 93.94%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 92.74%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 93.29%. Additionally, GRU only without GCN and feature extraction achieved an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 96.67%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 94.59%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 93.45%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 93.96%, indicating further gains. Also, Hybrid GCN-GRU without feature extraction and preprocessing models reached an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:acc{u}_{y}$$\end{document} of 97.22%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:pre{c}_{n}$$\end{document} of 95.17%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:rec{a}_{l}$$\end{document} of 94.21%, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F1}_{Score}$$\end{document} of 94.60%, highlighting slightly improved performance. Table 11Ablation of the DNAE-ADGC methodology.Model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MobileNetV3Large+MMWF (without CGBF)94.7892.6491.6391.82MobileNetV3Large+CGBF (without MMWF)95.3493.4092.2292.56GCN only (without GRU and feature extraction)95.9993.9492.7493.29GRU only (without GCN and feature extraction)96.6794.5993.4593.96Hybrid GCN-GRU (without feature extraction and preprocessing )97.2295.1794.2194.60DNAE-ADGC (Hybrid GCN-GRU+MobileNetV3Large with preprocessing techniques)97.9295.8794.9195.20 Table 11Ablation of the DNAE-ADGC methodology.Model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}1}_{\boldsymbol{S}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$$\end{document} MobileNetV3Large+MMWF (without CGBF)94.7892.6491.6391.82MobileNetV3Large+CGBF (without MMWF)95.3493.4092.2292.56GCN only (without GRU and feature extraction)95.9993.9492.7493.29GRU only (without GCN and feature extraction)96.6794.5993.4593.96Hybrid GCN-GRU (without feature extraction and preprocessing )97.2295.1794.2194.60DNAE-ADGC (Hybrid GCN-GRU+MobileNetV3Large with preprocessing techniques)97.9295.8794.9195.20 Ablation of the DNAE-ADGC methodology. The results in Table 12 demonstrates that the proposed DNAE-ADGC methodology attains grater performance on the SIPaKMeD database related to existing algorithms 38,39. It attains the maximum accuracy of 97.80%. These outcomes prove the robustness and generalization capability of the proposed method across different dataset. Table 12Comparison outcome of the DNAE-ADGC approach with other SIPaKMeD dataset.SIPaKMeD datasetMethodsAccuracyPrecisionRecallF1 ScoreCNN + LDA+kNN97.5490.8792.9890.714-layer CNN91.1390.1292.4590.35SVM96.8090.5092.1993.42CerCan-Net97.7092.1293.3991.95MSENet97.2193.9393.4693.54DNAE-ADGC97.8095.0194.1294.98 Table 12Comparison outcome of the DNAE-ADGC approach with other SIPaKMeD dataset.SIPaKMeD datasetMethodsAccuracyPrecisionRecallF1 ScoreCNN + LDA+kNN97.5490.8792.9890.714-layer CNN91.1390.1292.4590.35SVM96.8090.5092.1993.42CerCan-Net97.7092.1293.3991.95MSENet97.2193.9393.4693.54DNAE-ADGC97.8095.0194.1294.98 Comparison outcome of the DNAE-ADGC approach with other SIPaKMeD dataset. Figure 18 offers various Class Activation Map (CAM) visualizations that illustrate how the DL method focuses on significant regions in images for classification. The highlighted areas in red and yellow indicate regions with higher importance, where the model strongly attends to features relevant for prediction, while the blue regions represent less significant areas. These visualizations help in understanding the model’s decision-making process by revealing spatial patterns and feature localization. Fig. 18Class activation map (CAM) visualizations highlighting regions of interest in microscopic images. Class activation map (CAM) visualizations highlighting regions of interest in microscopic images. Conclusion This paper introduces a DNAE-ADGC methodology using medical imaging. The primary purpose of the paper is to empower gynaecological cancer diagnostics by developing an accurate, efficient, and intelligent detection framework using advanced techniques. Primarily, the image pre-processing phase deploys a two-level process named an adaptive filter. Followed by, the feature extraction process is executed by the MobileNetV3Large technique. Besides, the DNAE-ADGC model employs the GCN-GRU method to detect and classify gynaecological cancer. At last, the XAI model applies Grad-CAM to develop the transparency, interpretability, and reliability of AI models. The comparative outcome of the DNAE-ADGC methodology depicted a greater accuracy value of 97.92% over other models under the Malhari dataset. The proposed system enhances the interpretability, robustness and real-world applicability. The limitations include the reliance on specific dataset characteristics, which may affect generalizability across varied populations. Furthermore, the computational needs remain relatively high, limiting deployment on low-resource devices. Future work may focus on incorporating adaptive learning strategies, exploring multimodal data integration, and improving model explainability to improve usability and broader applicability in real-world medical settings. In the future work, DNAE-ADGC approach can be extended to stronger theoretical justification for the hybrid GCN-GRU algorithm, particularly for medical images. The method will also be tested on multi-institutional and larger databases to increase simplification and decrease Besides, the inclusion of the cross-dataset estimation in future drive be directed to certify clinical dependability. Additionally, the SMOTE and Focal loss techniques will be used to handle the class imbalance problem.