Explainable AI based hybrid DRM-Net transfer learning model for breast cancer detection and classification using ultrasound images Authors: Shahid Mohammad Ganie, Majid Bashir Malik, Mir Aadil, Ghazala Muteeb, Mohd Farhan, Mohammad Aatif Published: 2025 ABSTRACT Breast cancer is a serious health concern and one of the leading causes of cancer-related deaths among women worldwide. The early diagnosis of breast cancer is crucial for successful treatment and improved patient outcomes. Advanced computational techniques, particularly deep transfer learning have gained considerable attention for their effectiveness in medical image computing. Initially, six transfer learning models with different specifications are trained, validated and tested on the breast ultrasound image dataset. Additionally, a novel hybrid model, DRM-Net, is developed by stacking the top three TL models based on concatenation and flattening of deep dense layers. Various techniques, including image preprocessing, image masking, data augmentation, and hyperparameter tuning, are incorporated to enhance the overall performance of the considered models. The models are thoroughly evaluated using various standard metrics and statistical methods. Furthermore, to ensure transparent decision-making, the proposed model is interpreted using XAI-based class activation mapping (CAM) method with different versions. The proposed DRM-Net model demonstrated superior predictive performance compared to the other six transfer learning models. Among all the models, DRM-Net achieved the highest overall performance, with an accuracy of 96.71%, precision of 96%, recall of 97%, F1-score of 97%, and an AUC value of 99%, respectively. It also outperformed state-of-the-art similar studies in the literature. The experimental results demonstrate the potential utility of the hybrid DRM-Net model in improving the accuracy of breast cancer diagnosis, which could facilitate informed clinical decision-making. The proposed model can also be applied to other diseases where ultrasound imaging is available Introduction According to the World Health Organisation (WHO), Breast Cancer (BC) is one of the most common types of cancer and has affected millions of people globally (https://www.who.int/publications/i/item/who-position-paper-on-mammography-screening). In underdeveloped countries, BC rates are significantly higher than in developed countries, particularly among women. Each year, approximately 1.38 million individuals are diagnosed with breast cancer, and unfortunately, one-third of those diagnosed do not survive. The global death rate from cancer is 9.6 million, but there are also 1.7 million individuals who survive after receiving treatment for cancer1,2. The traditional method of classifying breast cancer, primarily based on clinicopathological characteristics and the analysis of common biomarkers, might not precisely reflect the varied clinical outcomes associated with each distinct subtype of breast cancer. Understanding the intricate biological processes involved in the onset and progression of cancer poses a significant challenge3. Several approaches exist for detecting breast cancer in its early stages, including undergoing diagnostic tests, observing signs of abnormality, or seeking advice from a local healthcare professional4. Implementing these strategies can lower the mortality rate and enhance the likelihood of successful treatment. Grasping the intricate biological mechanisms that contribute to the onset and progression of cancer poses a formidable challenge4,5. However, recent advances in high-throughput technologies have provided valuable insights into the genetic modifications and biological processes involved in breast cancer6. These discoveries offer novel perspectives on treatment strategies and patient classifications that can impact the management of individuals with breast cancer. Due to substantial healthcare expenses and elevated morbidity linked to cancer, researchers have been driven to create more precise models for cancer detection2. The primary challenges in recognizing and pinpointing cancerous cells in breast cancer images stem from their variations in size, shape, and position. Additionally, other anomalies like mastitis, adenopathy, and granuloma may coexist in breast images, further complicating the distinction of cancer cells5. Although conventional machine learning (ML) techniques are employed for cancer detection, their complex nature has resulted in reduced efficiency and accuracy in performance7,8. To overcome these challenges, deep learning (DL) has been introduced, demonstrating the capability to generate superior feature representations, thereby serving as a potent tool for addressing image classification and object localization tasks in cancer detection9. This paper provides an overview of recent research concerning classification of medical imaging data related to breast cancer, such as breast mammogram scans10. The present study scrutinizes and applies existing transfer learning-based AI applications for breast mammogram scans to underscore the potential of computer-aided diagnosis (CAD). Utilizing mammogram scan datasets, this research demonstrates that transfer learning in medical imaging enhances the accuracy, efficiency, and reliability of disease prediction within the healthcare system. Motivation In the identification of malignancy, computer-aided approaches outperformed radiologists. Image identification/classification and object recognition have become easier and more accurate as deep neural networks have advanced, particularly when it comes to medical imaging11. Multimodal imaging is playing an increasingly important role in improving early detection and streamlining the diagnostic process12. As powerful computers and extensive datasets become increasingly accessible, incorporating artificial intelligence (AI) into the current workflow of multimodality imaging for CAD diagnosis is a promising method for assisting radiologists in meeting the growing demand for imaging assessments and accurate early diagnosis13. Over the years, clinical management of breast cancer has included a range of therapeutic approaches, including chemotherapy, targeted therapies, and immunotherapy5. In clinical settings, different imaging modalities, including computed tomography (CT), nuclear medicine imaging, X-rays, magnetic resonance imaging (MRI), etc., are utilized for breast cancer diagnosis14. Their interpretation requires highly trained radiologists due to the complexity of the images and potential overlap with other breast abnormalities15. Figure 1 depicts the top 6 cancers with their incidence rates and death rates (https://www.who.int/news-room/fact-sheets/detail/cancer).Fig. 1Comparison of incident cases and death rates of cancer types. Comparison of incident cases and death rates of cancer types. As artificial intelligence has advanced, ML techniques have been explored in various sectors, with a notable impact in healthcare for tasks as disease detection16, prediction17, diagnosis18, and prognosis19. Also, DL has demonstrated substantial results and is now widely employed in medical image analysis20,21. These models are capable of identifying complex patterns within high-dimensional data, surpassing the limitations of traditional image interpretation methods that typically rely on expert visual assessment22–24. Convolutional neural networks (CNNs) are effective in extracting the local and global level features efficiently from image morphology25, contributing to improved accuracy in classification and detection tasks26,27. In recent years, TL models have produced desirable outcomes, especially when applied to large-scale medical imaging datasets and have emerged as powerful techniques for enhancing disease diagnostic performance28. Why deep transfer learning Transfer learning models have gained significant attention and are widely implemented architectures in various applications, including medical image diagnosis29, computer vision30, natural language processing31, speech recognition32, and many more33. TL involves using pre-trained neural networks to transfer the insights and knowledge acquired from one problem statement to another across similar domains34. This is typically achieved by employing a CNN pre-trained on a large-scale image dataset, which is subsequently fine-tuned on the target breast ultrasound image dataset. Through this fine-tuning process, the model learns domain-specific features relevant to breast cancer detection and classification without the need for extensive labelled data. These models have demonstrated great potential by producing reliable outcomes with minimal computational resources, thereby enhancing the efficiency of disease detection and classification. The principal advantages of integrating TL models in breast cancer detection and classification are illustrated in Fig. 2.Fig. 2Advantages of transfer learning for breast cancer detection. Advantages of transfer learning for breast cancer detection. Contribution of the paper In this study, an extensive assessment of the considered TL models is conducted using ultrasound images to propose a hybrid DRM-Net model to enhance the prediction capability of pre-trained models for breast cancer detection and classification. The primary contributions of this research are as follows:The original considered image dataset is resized (512 × 512 × 3), preprocessed, masked, and augmented to enhance experimental usability.Six diverse transfer learning models, viz. VGG16, EfficientNetB0, InceptionV3, DenseNet169, MobileNet, and ResNet50 are trained, validated and tested with the processed dataset.A novel transfer learning model (DRM-Net) is developed by stacking DenseNet169, ResNet50, and MobileNet. The outputs of all three models are individually flattened and subsequently concatenated.A comprehensive evaluation is conducted using various performance metrics and statistical analyses to assess the predictive significance of the considered models and the proposed DRM-Net model.The DRM-Net model achieved higher accuracy (96.71%), precision (96%), recall (97%), F1-score (97%), and AUC (99%), showing approximately ~ 6% higher accuracy and ~ 5–7% higher in other metrics compared to the TL models.DRM-Net model interpretability is enhanced using Explainable AI techniques, incorporating different versions of the class activation mapping (CAM) method.The proposed DRM-Net model is rigorously validated and compared with the considered transfer learning (TL) techniques and other state-of-the-art work. The original considered image dataset is resized (512 × 512 × 3), preprocessed, masked, and augmented to enhance experimental usability. The original considered image dataset is resized (512 × 512 × 3), preprocessed, masked, and augmented to enhance experimental usability. Six diverse transfer learning models, viz. VGG16, EfficientNetB0, InceptionV3, DenseNet169, MobileNet, and ResNet50 are trained, validated and tested with the processed dataset. Six diverse transfer learning models, viz. VGG16, EfficientNetB0, InceptionV3, DenseNet169, MobileNet, and ResNet50 are trained, validated and tested with the processed dataset. A novel transfer learning model (DRM-Net) is developed by stacking DenseNet169, ResNet50, and MobileNet. The outputs of all three models are individually flattened and subsequently concatenated. A novel transfer learning model (DRM-Net) is developed by stacking DenseNet169, ResNet50, and MobileNet. The outputs of all three models are individually flattened and subsequently concatenated. A comprehensive evaluation is conducted using various performance metrics and statistical analyses to assess the predictive significance of the considered models and the proposed DRM-Net model. A comprehensive evaluation is conducted using various performance metrics and statistical analyses to assess the predictive significance of the considered models and the proposed DRM-Net model. The DRM-Net model achieved higher accuracy (96.71%), precision (96%), recall (97%), F1-score (97%), and AUC (99%), showing approximately ~ 6% higher accuracy and ~ 5–7% higher in other metrics compared to the TL models. The DRM-Net model achieved higher accuracy (96.71%), precision (96%), recall (97%), F1-score (97%), and AUC (99%), showing approximately ~ 6% higher accuracy and ~ 5–7% higher in other metrics compared to the TL models. DRM-Net model interpretability is enhanced using Explainable AI techniques, incorporating different versions of the class activation mapping (CAM) method. DRM-Net model interpretability is enhanced using Explainable AI techniques, incorporating different versions of the class activation mapping (CAM) method. The proposed DRM-Net model is rigorously validated and compared with the considered transfer learning (TL) techniques and other state-of-the-art work. The proposed DRM-Net model is rigorously validated and compared with the considered transfer learning (TL) techniques and other state-of-the-art work. Organization of the paper The organization of the paper is shown in Fig. 3. It outlines the structure of the paper, highlights the flow and key components discussed in subsequent sections.Fig. 3Structure of the paper. Structure of the paper. Related work Over the past few years, computer-aided diagnosis systems have shown promise in detecting and predicting breast cancer using medical imaging. Various DL and TL models have been utilized to automate breast cancer diagnosis, significantly improving diagnostic precision compared to traditional methods35. Extensive research has been conducted encompassing various datasets, algorithms, and methodologies for breast cancer detection and classification. This section reviews some similar studies that have applied DL and TL techniques for breast cancer detection and classification. Lu et al.36 proposed a novel computer-aided diagnosis system for breast cancer detection called BCDNet. They utilized deep Convolutional Neural Network (CNN) models for extracting image representations and introduced an adaptive backbone selection algorithm to identify the most suitable CNN for the task. Among all the models, backbone BCDNet achieved the highest accuracy, sensitivity, precision, and F1-score of 93.97%, 95.24%, 97.68%, and 96.42%, respectively. Uysal and Kose37 conducted research to explore the application of pre-trained deep learning models for classifying breast cancer using ultrasound images. The study aimed to differentiate between benign (non-cancerous), malignant (cancerous), and normal samples. Three models were utilised, namely ResNet50, ResNeXt50, and VGG16. The results showed that ResNeXt50 had the highest accuracy score of 85.83%, outperforming the other models. However, ResNet50 demonstrated greater stability and better class-based results. Li et al.38 presented a deep learning model called BUSnet for detecting breast tumor lesions in ultrasound images. The employed model has a two-stage approach that uses unsupervised region proposal and bounding-box regression algorithms. Additionally, a post-processing technique was introduced to improve the performance of the disease detection results. The results indicated that BUSnet achieved an accuracy of 0.651, precision of 0.832, recall of 1.000, and f1-score of 0.789. Gu et al.39 implemented a deep learning model that can differentiate between benign and malignant breast lesions in ultrasound images using a large dataset collected from multiple centers of hospitals in China. The model demonstrated positive outcomes on the test sets, achieving an AUC of 0.908, a sensitivity of 83.23%, a specificity of 83.61%, and an accuracy of 83.48% for the internal test set. Additionally, for the external test sets, the model attained an AUC of 0.913, a sensitivity of 88.84%, a specificity of 83.77%, and an accuracy of 86.40%. Anand et al.40 developed a classification model using transfer learning for breast cancer detection based on the CBIS-DDSM dataset. In their work, the authors have explored the learning properties of layers for CNNs to fine-tune the model, achieving an accuracy of 75% and reducing false-negative rates, which indicates an improvement over existing methods. Hijab et al.41 proposed a deep learning model for the classification of ultrasound images to detect breast cancer. Initially, the baseline approach was employed to train the basic architecture of CNN, and a further pre-trained, fine-tuned VGG16 model was used to tackle the overfitting problem. As per the experiment, the model achieved an accuracy rate of 97% and an AUC of 98%. The authors suggested that this work can be used to explore the different imaging modalities for better prediction and detection of the disease. Rana et al.42 introduced a framework for breast cancer utilizing transfer learning models, primarily based on histopathological images. In this work, the authors have discussed and presented the effectiveness of transfer learning models without preprocessing techniques for better classification of tumors. Various transfer learning models were utilized, including LENET, VGG16, DarkNet53, DarkNet19, ResNet50, Inception, and Xception. Initially, Xception computed the best accuracy of 83.07% using the unbalanced dataset, while DarkNet53 attained the best accuracy rate of 87.17% by adding a new parameter, namely Balanced Accuracy (BAC). Shelke et al.43 implemented a model using transfer learning for the detection and classification of breast cancer based on the breast ultrasound scanning imaging (BUSI) dataset. The authors used pre-trained models including Sequential, Xception, DenseNet201, VGG16, and InceptionResNetV2, and compared them with the ensemble models. The EL model achieved the highest accuracy rate of 87.90%. Alruily et al.44 presented a novel approach using Vision Transformer (ViT) models for classification of benign and malignant breast tumors. The proposed ViT model achieved a high accuracy of 94.49% and an AUC of 0.921 for breast cancer prediction. The model outperforms traditional CNNs like VGG16 and DenseNet, demonstrating superior generalization without requiring data augmentation. The authors have suggested that using XAI methods like Grad-CAM, ViT’s self-attention mechanism inherently enhances interpretability for breast cancer diagnosis. Wang et al.45 proposed an anchor-free deep learning approach (FCOS) for detecting and classifying benign and malignant breast lesions in ultrasound images. To improve the image contrast, a segmentation-based enhancement (SBE) technique is used for better lesion detection. The proposed method achieves a mean average precision (mAP) of 90.2%, outperforming YOLO V3 (anchor-based method) and demonstrating high recall (98% for benign, 96.3% for malignant lesions), ensuring fewer missed cancer cases. The authors suggested that use of XAI can help in the segmentation step, inherently enhancing interpretability by localizing tumor regions. Daoud et al.46 implemented a hybrid model for breast tumor classification. The study employs a feature selection algorithm (VGG19) to identify the most relevant tumor characteristics, achieving a high accuracy of 96.10% when combining deep and morphological features. The proposed method is validated using cross-validation on 380 images and an independent external dataset of 163 images, demonstrating strong generalization. Overall, these studies demonstrate that DL and TL have the potential to enhance the effectiveness of medical imaging data in classification and detection of breast cancer. Also, the adoption of pre-trained TL networks requires less computational resources and training time, as they are employed on large image datasets to transfer knowledge substantially across multiple applications. However, most existing studies primarily focus on independent transfer learning without exploring advanced hybrid and ensemble approaches. Additionally, the use of XAI techniques for feature identification and importance remains underexplored, limiting model interpretability and explainability. There is an opportunity for further research to identify the optimal architecture for TL and the most effective fine-tuning strategies for breast cancer detection. Furthermore, prior research has largely emphasized accuracy without conducting rigorous statistical validation to assess the significance of model performance. Our study addresses these gaps by proposing a hybrid deep TL model designed to improve breast cancer diagnosis using ultrasound imaging. The inclusion of diverse TL models from different families with varied configurations enabled the assessment of individual model performance to design a novel hybrid DMI-Net architecture. Also, the hybrid preprocessing techniques, including image resizing, augmentation, and masking, are utilized to enhance the quality of the images and experimental usability. Model performance is assessed using standard evaluation metrics and further validated through statistical methods, including the Friedman aligned ranks test and post hoc analysis. We further enhance model interpretability using XAI methods, including Grad-CAM, Grad-CAM + +, and Score-CAM, to highlight and identify key features for breast cancer prediction. The proposed model is compared with considered TL models and similar recent research works, highlighting the improved performance and effectiveness of the proposed approach. Research methodology The following section explains the comprehensive materials and methodology employed in this study for the detection of breast cancer. It details the experimental requirements and procedural steps carried out in this paper. The section also provides an overview of the dataset utilized, followed by a description of the image preprocessing techniques, including resizing, augmentation, and masking. Additionally, an in-depth discussion of the selected TL models is given subsequent subsection. Framework The proposed model is illustrated in Fig. 4. Following the acquisition of breast ultrasound images, preprocessing and augmentation were performed to ensure experiment suitability. The processed dataset was then split into training, validation, and testing sets. Eight transfer learning models were employed based on this data. The top three among them were chosen and combined to construct a novel prediction model. The model underwent iterative fine-tuning to enhance classification accuracy while minimizing the necessary training time. Training and validation were conducted to categorize two cancer classes and a normal class. Subsequently, the model underwent testing.Fig. 4Framework of the proposed methodology. Framework of the proposed methodology. Dataset description Using computational techniques on breast mammogram scans has shown promising results in breast cancer classification, detection, and diagnosis. A publicly available Breast Ultrasound Images Dataset (https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset) with PNG-format images of an average size of 500 × 500 pixels has been utilised to develop a TL model for breast cancer detection. The dataset contains data from 650 female patients and is separated into three categories: normal, benign, and malignant, with the number of images in each category listed in Table 1. Also, Fig. 5 shows examples of the data samples.Table 1Detailed sample-wise distribution of dataset before resampling.DatasetTotal imagesImage propertiesBenignMalignantNormalUltrasound1578No. of images891421266ImagesClass123DatasetLabel012Fig. 5Sample images from breast ultrasound images dataset (a) benign, (b) malignant, and (c) normal. Table 1Detailed sample-wise distribution of dataset before resampling.DatasetTotal imagesImage propertiesBenignMalignantNormalUltrasound1578No. of images891421266ImagesClass123DatasetLabel012 Detailed sample-wise distribution of dataset before resampling. Sample images from breast ultrasound images dataset (a) benign, (b) malignant, and (c) normal. Benign Benign (https://my.clevelandclinic.org/health/diseases/6270-benign-breast-disease) breast conditions elevate the likelihood of developing breast lumps. The discovery of a lump can be alarming; however, it’s important to note that these alterations in the breast are noncancerous. Some forms of breast disease heighten the risk of developing breast cancer. It is crucial to inform your healthcare provider if you notice any breast lumps or changes. Malignant Malignant (https://www.verywellhealth.com/malignant-breast-disease-429987) tumors undergo assessment and classification using a specified system that is structured to determine their severity. The presence of a malignant tumor indicates the presence of breast cancer or another type of cancer. Malignant tumors can exhibit aggressiveness and have the potential to spread to neighbouring tissues. Treatment is imperative for addressing and managing malignant tumors. Normal The last folder is the normal ultrasound images, which are noncancerous only. Image preprocessing Image preprocessing is a critical step in improving both image quality and the effectiveness of algorithms used in the development of computer-aided diagnosis (CAD) systems47. To ensure reliability, the selected transfer learning models in this study are trained, tested, and validated using preprocessed data. A hybrid preprocessing approach is adopted, comprising image resizing (to match model-specific input dimensions), data augmentation (to address class imbalance), and masking (to enhance feature learning). Each of the preprocessing steps is discussed in detail in the subsequent subsections. Image resizing The image resizing is an important preprocessing step required by DL models to ensure the consistency in input image dimensions and can significantly impact training time and performance. The actual datasets utilized in this work comprise images of varying dimensions with an average image size of 500*500 pixels both within and across classes. All the images are normalized and resized to 512 × 512, a uniform dimension to ensure preservation of image features and compatibility with the model architectures. The classes undergo label encoding, i.e., 0 for class benign, 1 for class malignant, and 2 for class normal, respectively. This label encoding facilitates efficient classification during the model training process. Data augmentation The dataset used for the experiment was imbalanced across the classes, which can negatively impact the performance of the model during the training phase. To address this issue, an oversampling technique was applied by augmenting the number of images in the minority classes. Initially, the dataset comprised 1578 images, categorized into benign, malignant, and normal as 891, 421, and 266 images, respectively. To generate synthetic variants of the existing images, data augmentation was applied with different particulars (shearing, rotation, flipping, and zooming) and their corresponding values to ensure the generation of varied yet structurally consistent images. Following augmentation, the final dataset used for experimentation contains 2531 images with class distributions of benign, malignant, and normal as 891, 842, and 798 images, as shown in Table 2.Table 2Detailed sample-wise distribution of dataset after resampling.Total imagesImage propertiesBenignMalignantNormal2531No. of images891842798Class123Label012 Table 2Detailed sample-wise distribution of dataset after resampling.Total imagesImage propertiesBenignMalignantNormal2531No. of images891842798Class123Label012 Detailed sample-wise distribution of dataset after resampling. Image masking is a technique used in image processing, especially in medical domain, to identify, hide or highlight specific parts of an image48. It involves applying a binary or multi-channel mask to an image to filter out irrelevant information while keeping the important features. In our study, each image was resized to 512 × 512 pixels and then cropped to 224 × 224 pixels to focus on the central area where lesions are usually found. Gaussian filtering was applied to improve image quality and reduce noise, which gently smooths the images, and Sobel filtering to highlight edges. Also, the pixel values were normalized using a mean of [0.485, 0.456, 0.406] and a standard deviation of [0.229, 0.224, 0.225] to match the considered pre-trained transfer learning models’ requirements, ensuring consistent and high-quality input data for accurate breast cancer classification. The steps for processing of image masking, including blurring the image, edge detection, and histogram equalization, are described and calculated using Eq. 1, Eq. 2, and Eq. 2, respectively. In addition, Fig. 6 shows the actual image, mask, and overlayed image of each sample of classes of the considered dataset.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{blurred}={I}^{*}G(\upsigma )$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E=\nabla I$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{equalized}=HistEqual(I)$$\end{document}where, in Eq. 1, Gaussian blurring applied to an image I, G(σ) is a gaussian filter with standard deviation σ, and the operation ∗ denotes convolution between the image and the Gaussian kernel. Similarly, in Eq. 2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\nabla I$$\end{document} represents the gradient of the image. Furthermore, in Eq. 3, HistEqual(I) represents histogram equalization, a technique used to enhance contrast.Fig. 6Sample wise masking and overlayed dataset after resampling. Sample wise masking and overlayed dataset after resampling. Transfer learning models TL models have gained considerable attention and have demonstrated significant potential in the healthcare industry, particularly in medical image computing34. The TL pre-trained architectures can extract the local and global features from diverse images efficiently. The models are trained over large image datasets like ImageNet, enabling efficient knowledge transfer for medical image computing49. The subsection outlines the considered TL models utilized in this study for breast cancer detection and classification. VGG16 The VGG1650 is a classical CNN architecture consisting of 16 weight layers, which include 13 convolutional layers and 3 fully connected layers. After every two or three convolutional layers, max-pooling layers are applied to reduce the spatial dimensions of feature maps. Max-pooling layers are inserted after every two or three convolutional layers to reduce the spatial dimensions of the feature maps. This architecture has been widely adopted across various applications, particularly in computer vision tasks such as image classification, object detection, and segmentation, consistently demonstrating strong performance. A key advantage of VGG16 architecture is that it uses very small convolution filters (3 × 3) in all convolutional layers, enabling the construction of deeper networks with a manageable number of parameters. The output from CNN layers of architecture is calculated using Eq. 4.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=f({W}^{*}x+b)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document} represents the input image, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W$$\end{document} is the convolutional layer’s weight matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document} denotes the bias, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f$$\end{document} is the activation function. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y$$\end{document} represents the output of the feature map, enabling relevant and hierarchical feature extraction. EfficientNetB0 EfficientNetB051 is a baseline CNN architecture designed through neural architecture search (NAS). These models are designed to achieve high performance while minimizing computational costs. In this architecture, a new scaling method is introduced that scales depth, width, and resolution using static scaling coefficients, ensuring better performance and efficacy. The EfficientNet architecture is comprised of 16 blocks, where each block is determined by the number of channels in its CNN layers. The compound scaling in EfficientNet architecture is mathematically expressed in Eqs. 5, 6, and 7.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d={\alpha }^{\varnothing }$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w={\beta }^{\varnothing }$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r={\gamma }^{\varnothing }$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w$$\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}$$r$$\end{document} represent the depth (number of layers), width (number of channels per layer) and resolution (input image size), respectively. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\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}$$\gamma$$\end{document} are constant scaling factors determined through grid search. Also, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varnothing$$\end{document} is the compound coefficient that controls model scaling. InceptionV3 InceptionV352 is a widely popular and recognized CNN architecture that incorporates several key design principles, such as convolutional layer, depth-wise separable convolutions, residual connections, and inception modules. This architecture introduces the swish activation function, which improves upon traditional activation functions by enabling smoother gradient flow and better convergence. At its core, InceptionV3 is built upon the inception module, which effectively separates the cross-channel correlations from spatial relationships within CNN feature maps. The Inception block of this architecture is mathematically expressed in Eq. 8.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=Concat \left({Conv}_{1\times 1}\left(x\right),{Conv}_{3\times 3}\left({Conv}_{1\times 1}\left(x\right)\right), {Conv}_{5\times 5}\left({Conv}_{1\times 1}\left(x\right)\right), {Conv}_{1\times 1}\left({MaxPool}_{3\times 3}\left(x\right)\right)\right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Concat$$\end{document} denotes the concatenation of feature maps along the channel axis, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Conv$$\end{document} is a convolution operation with ReLU or batch normalization, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MaxPool$$\end{document} represents the down sampling to reduce spatial dimensions. DenseNet169 DenseNet12153 is a deep CNN comprising 169 layers, structured according to the DenseNet architecture. In this model, each layer is connected directly to every other succeeding layer in a feedforward manner, significantly enhancing information flow and gradient propagation across the network. Such dense connectivity encourages feature reuse, reduces overall parameter count, and effectively addresses the vanishing gradient issue often encountered in deep CNNs. The outputs in the DenseNet169 architecture for connected and transition layers are calculated using Eqs. 9 and 10.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{i}=f({H}_{0},{H}_{1},{H}_{2},{H}_{3}, \dots \dots ..,{H}_{i-1})$$\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}$${H}_{i+1}=f(BN ({W}_{i+1}\times {H}_{i}))$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{i}$$\end{document} denotes the current layer output, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f$$\end{document} is the 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}$$({H}_{0},{H}_{1},{H}_{2},{H}_{3}, \dots ,{H}_{i-1})$$\end{document} represents the concatenated outputs of all preceding layers. Also, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H}_{i+1}$$\end{document} corresponds to the output after applying the weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{i+1}$$\end{document} in the convolutional layer, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BN$$\end{document} indicating batch normalization. MobileNet MobileNet54 is a widely used deep neural network architecture drafted for mobile and embedded devices characterized by limited computational resources. The architecture is founded on a lightweight building block known as a MobileNet unit, comprising a depthwise separable convolution layer succeeded by a pointwise convolution layer. The depthwise separable convolution is a factorised convolution that decomposes a standard convolution into a depthwise convolution and a pointwise convolution, which reduces the number of parameters and computations. The output of a MobileNet unit and inverted residual block can be calculated using Eqs. 11–15.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=BN(\sigma ({Conv}_{1*1}(DW(X))))$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{in}=X$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X=BN(\sigma ({Conv}_{1*1}(DW(X))))$$\end{document}14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X=BN(\sigma ({Conv}_{1*1}(X)))$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y={X}_{in}+X$$\end{document}where X is the input tensor, DW is the depthwise convolution operation, Conv1x1 is the pointwise convolution operation, σ is the activation function, BN is the batch normalisation operation, and Y is the output tensor. Also, Xin is the input tensor, X is the output tensor of the bottleneck layer, Conv1x1 and DW are the pointwise and depthwise convolution operations. ResNet50 ResNet5055 is a widely used deep CNN architecture across multiple computer vision applications. This network is designed to mitigate the vanishing gradient problem in deep networks. A new method, skip connections, is introduced, which is also known as residual connections, that helps to integrate the output of one layer with the previous layer’s output. It is comprised of one convolution + max pooling layer,16 residual blocks (bottleneck architecture), and final global average pooling + fully connected layer. Mathematically, the residual block can be expressed as given by Eq. 16.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y=F(x, \left\{{W}_{i}\right\}+x)$$\end{document}where x is denotes the input to the block, F represents the set of convolutional layers with weights Wi, and y represents the block output. The input x to the output y is added by skip connection to produce the final output of the block. Proposed DRM-Net model To identify the best-performing models among those discussed in the previous subsection (3.4), each model was individually trained, validated and tested using standard performance metrics. As far as our results are concerned, MobileNet, DenseNet169, and ResNet50 consistently demonstrated superior performance across most of the evaluation metrics. It can be observed that InceptionV3 and EfficientNetB0 outperformed in certain metrics, but ResNet50 was selected because it produced better testing accuracy and precision, which is crucial for breast cancer detection and classification. Therefore, based on this comparative analysis, the proposed DRM-Net model was developed by stacking DenseNet169, ResNet50, and MobileNet. The complete algorithm outlining this procedure is presented in Algorithm 1. Algorithm 1Developing the DRM-Net model. Developing the DRM-Net model. Model architecture The architecture of the proposed DRM-Net model is illustrated in Fig. 7. The network takes an input image of shape 512 × 512 × 3 and produces prediction outputs across three classes: benign, malignant, and normal. The model integrates three stacked TL architectures, including DenseNet169, ResNet50, and MobileNet, each followed by dedicated dense layers. Subsequently, identical convolutional layers of size 7 × 7 × 1024 are applied to each model, followed by individual max-pooling layers to downsample the feature maps. Each pathway’s output is then flattened and passed through a sequence of three fully connected layers (with sizes 1024, 512, and 256, respectively). The resulting feature representations from all three branches are concatenated, and the final classification is determined using a majority voting mechanism. This fusion enables a more robust and comprehensive representation of extracted features, leveraging the strengths of all three TL models. The architectural description of DRM-Net is shown in Table 3.Fig. 7Proposed hybrid DRM-Net’s architecture.Table 3Description of the DRM-Net model’s architecture.LayerDescription and interpretationInput Layer (512 × 512 × 3)This layer accepts high-resolution breast cancer images to enhance the proposed model’s ability to effectively differentiate between benign, malignant, and normal images in considered datasetDenseNet169, ResNet50, and MobileNetThese pretrained top-performing models are used for extracting different features from images to build the hybrid DRM-Net modelConcatenation LayerThis layer merges the outputs of MobileNetV2, ResNet50, and VGG16 and allows for a comprehensive feature representation to build the DRM-Net model for better prediction of breast cancerBatch normalizationIt helps to normalize the activations within the network to improve stability and convergence and to reduce internal covariate shift during training of the considered models for breast cancerConv2D + MaxPoolingIt enhances the DRM-Net model’s ability to detect fine-grained differences, such as class 1: irregular cell growth, but not fully invasive, class 2: aggressive, deeply invasive cancer patterns, and class 3: normal tissue with smooth patternsFlatten LayerIt converts the high-dimensional feature maps into a 1D feature vector. This transition from convolutional layers to dense layers ensures that important features are retained and properly classified across the samples of imagesDropout LayersIt helps the DRM-Net model learn only the most significant features and ensures it does not memorize noise but instead focuses on distinguishing characteristics of cancerous and non-cancerous tissues of breast imagesDense LayersThese layers take the extracted features from convolutional layers and make the final decision on classifying breast cancer images into three categories (benign, malignant and normal)Final Output Layer (3 neurons)It uses softmax activation to predict the probability of each class, such as benign, malignant and normal, of breast cancer images Proposed hybrid DRM-Net’s architecture. Table 3Description of the DRM-Net model’s architecture.LayerDescription and interpretationInput Layer (512 × 512 × 3)This layer accepts high-resolution breast cancer images to enhance the proposed model’s ability to effectively differentiate between benign, malignant, and normal images in considered datasetDenseNet169, ResNet50, and MobileNetThese pretrained top-performing models are used for extracting different features from images to build the hybrid DRM-Net modelConcatenation LayerThis layer merges the outputs of MobileNetV2, ResNet50, and VGG16 and allows for a comprehensive feature representation to build the DRM-Net model for better prediction of breast cancerBatch normalizationIt helps to normalize the activations within the network to improve stability and convergence and to reduce internal covariate shift during training of the considered models for breast cancerConv2D + MaxPoolingIt enhances the DRM-Net model’s ability to detect fine-grained differences, such as class 1: irregular cell growth, but not fully invasive, class 2: aggressive, deeply invasive cancer patterns, and class 3: normal tissue with smooth patternsFlatten LayerIt converts the high-dimensional feature maps into a 1D feature vector. This transition from convolutional layers to dense layers ensures that important features are retained and properly classified across the samples of imagesDropout LayersIt helps the DRM-Net model learn only the most significant features and ensures it does not memorize noise but instead focuses on distinguishing characteristics of cancerous and non-cancerous tissues of breast imagesDense LayersThese layers take the extracted features from convolutional layers and make the final decision on classifying breast cancer images into three categories (benign, malignant and normal)Final Output Layer (3 neurons)It uses softmax activation to predict the probability of each class, such as benign, malignant and normal, of breast cancer images Description of the DRM-Net model’s architecture. Model hyperparameters and augmentation The details of hyperparameter settings and data augmentation for DRM-Net are listed in Table 4. The selected hyperparameters, including optimizers, learning rate, loss function, dropout probability, batch size, and number of epochs, ensure efficient training and prevent overfitting. Additionally, various data augmentation tasks such as shear, zoom, rotation, and flipping are applied to enhance the model’s ability to generalize across diverse breast cancer image variations.Table 4Hyperparameter settings and data augmentation details of DRM-Net model.Tuning hyperparametersData augmentationHyperparameterValueParticularValueOptimiser 1AdamShear range0.2Optimiser 2RMSpropZoom range0.2Learning rate0.001Rotation range24LossCategorical cross-entropyVertical flipTrueDropout probability0.3Horizontal flipTrueBatch size32Vertical shearing0.2Number of epochs30Horizontal shearing0.2 Table 4Hyperparameter settings and data augmentation details of DRM-Net model.Tuning hyperparametersData augmentationHyperparameterValueParticularValueOptimiser 1AdamShear range0.2Optimiser 2RMSpropZoom range0.2Learning rate0.001Rotation range24LossCategorical cross-entropyVertical flipTrueDropout probability0.3Horizontal flipTrueBatch size32Vertical shearing0.2Number of epochs30Horizontal shearing0.2 Hyperparameter settings and data augmentation details of DRM-Net model. Experiment, results and performance analysis In this section, the experimental setup, target outcomes, and performance evaluation of the proposed approach are comprehensively discussed. The hardware and software configurations, as well as the evaluation metrics employed, are detailed in the subsequent subsections. Finally, the implementation of the proposed DRM-Net model and its comparative performance analysis are thoroughly presented to demonstrate its effectiveness against other baseline models. Hardware and software requirements The experiments were carried out on a Dell workstation operating in a Microsoft Windows environment. Python programming language was utilized within the Anaconda framework for model development and evaluation. The detailed specifications of the system configuration used in this study are provided in Table 5.Table 5System’s hardware and software specifications for the experiment.ComponentSpecificationProcessorIntel Core i9-10900 K with a base clock speed of 3.70 GHzGPUNVIDIA GeForce RTX 2080 TI featuring 11 GB GDDR6 memoryRAM64 GB DDR4, optimized for high-performance computingSSD500 GB NVMe SSD for rapid data access and storageHard Disk2 TB HDD for additional storage needsOperating SystemWindows 11 Pro for a secure and efficient user experiencePlatform (IDE)Anaconda with Jupyter Notebook for interactive coding and analysisProgramming LanguagePython, chosen for its extensive libraries and packages for DL models Table 5System’s hardware and software specifications for the experiment.ComponentSpecificationProcessorIntel Core i9-10900 K with a base clock speed of 3.70 GHzGPUNVIDIA GeForce RTX 2080 TI featuring 11 GB GDDR6 memoryRAM64 GB DDR4, optimized for high-performance computingSSD500 GB NVMe SSD for rapid data access and storageHard Disk2 TB HDD for additional storage needsOperating SystemWindows 11 Pro for a secure and efficient user experiencePlatform (IDE)Anaconda with Jupyter Notebook for interactive coding and analysisProgramming LanguagePython, chosen for its extensive libraries and packages for DL models System’s hardware and software specifications for the experiment. Performance evaluation metrics Evaluation metrics are used to assess the performance of a model on a problem statement. Different evaluation metrics are used depending on the problem type and the data’s nature. In this study, the experimental findings for the presented models are evaluated using various performance metrics, summarised in Table 6.Table 6Performance evaluation metrics.MetricsCalculationSignificance and interpretationAccuracy\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP + TN}{{TP + TN + FP + FN}}$$\end{document}Accuracy gives the classification success of the model, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TP,\;TN,\;FP\;{\text{and}}\;FN$$\end{document} denote true positives, true negatives, false positives, and false negatives. For breast cancer prediction, accuracy indicates that the model correctly identifies benign, malignant, and normal classesPrecision\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP}{{TP + FP}}$$\end{document}Precision measures the ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} to the sum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} and F \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Ps$$\end{document}, indicating how many positive predictions were actually correct. In breast cancer prediction, high precision means the model’s cancer class predictions are highly likely to be correct, which is crucial in medical diagnostics to minimize unnecessary testingRecall\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP}{{TP + FN}}$$\end{document}Recall measures the ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} to the sum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} and false negatives, showing how many positive cases were correctly identified. For breast cancer prediction, high recall indicates successful identification of cancer cases, critical for early detection, though it may reduce precisionF1-score\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{2 \times TP}{{2 \times TP + FP + FN}}$$\end{document}F1-score, the harmonic mean of precision and recall, indicates model effectiveness by balancing precision and recall in breast cancer stage prediction while maintaining low false positivesLoss\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{n}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times \mathop \sum \limits_{i = 1}^{n} \mathop \sum \limits_{c = 1}^{k}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {y_{ic} \times \log \left( {\hat{y}_{ic} } \right)} \right]$$\end{document} Loss measures discrepancy between prediction probabilities and target values for model evaluation. For breast cancer, \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} represents image count, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} denotes classes (3), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{ij}$$\end{document} is the actual label (one-hot encoded), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{ij}$$\end{document} is predicted class probability. Lower loss indicates better classification of breast cancer stagesMacro average\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{3}\mathop \sum \limits_{c = 0}^{2} A_{c}^{m}$$\end{document}Micro average computes the arithmetic mean of class-wise values for precision, recall, and f1-score. For breast cancer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document} represents classes and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m$$\end{document} denotes precision, recall or F1-score, ensuring balanced performance measurementWeighted average\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathop \sum \limits_{c = 0}^{2} w_{c}^{m}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times \frac{1}{3}\mathop \sum \limits_{c = 0}^{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{c}^{m}$$\end{document} Weighted loss is the mean of precision, recall, and F1-score metrics for each class, multiplied by class weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{0} + w_{1} + w_{2} + w_{3} = 1$$\end{document}. For breast cancer, it weights all classes equally and handles rare stages reliablyStandard deviation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sqrt {\frac{{\mathop \sum \nolimits_{c = 0}^{2} (S_{c}^{m} - \frac{1}{3}\mathop \sum \nolimits_{c = 0}^{2} A_{c}^{m} )^{2} }}{n}}$$\end{document}Standard Deviation measures how evaluation metrics like precision, recall, and F1-score deviate from their means across classes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=3$$\end{document}). High deviation indicates variable performance, while low deviation shows consistency across breast cancer stages Table 6Performance evaluation metrics.MetricsCalculationSignificance and interpretationAccuracy\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP + TN}{{TP + TN + FP + FN}}$$\end{document}Accuracy gives the classification success of the model, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TP,\;TN,\;FP\;{\text{and}}\;FN$$\end{document} denote true positives, true negatives, false positives, and false negatives. For breast cancer prediction, accuracy indicates that the model correctly identifies benign, malignant, and normal classesPrecision\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP}{{TP + FP}}$$\end{document}Precision measures the ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} to the sum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} and F \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Ps$$\end{document}, indicating how many positive predictions were actually correct. In breast cancer prediction, high precision means the model’s cancer class predictions are highly likely to be correct, which is crucial in medical diagnostics to minimize unnecessary testingRecall\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{TP}{{TP + FN}}$$\end{document}Recall measures the ratio of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} to the sum of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPs$$\end{document} and false negatives, showing how many positive cases were correctly identified. For breast cancer prediction, high recall indicates successful identification of cancer cases, critical for early detection, though it may reduce precisionF1-score\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{2 \times TP}{{2 \times TP + FP + FN}}$$\end{document}F1-score, the harmonic mean of precision and recall, indicates model effectiveness by balancing precision and recall in breast cancer stage prediction while maintaining low false positivesLoss\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{n}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times \mathop \sum \limits_{i = 1}^{n} \mathop \sum \limits_{c = 1}^{k}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {y_{ic} \times \log \left( {\hat{y}_{ic} } \right)} \right]$$\end{document} Loss measures discrepancy between prediction probabilities and target values for model evaluation. For breast cancer, \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} represents image count, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} denotes classes (3), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y}_{ij}$$\end{document} is the actual label (one-hot encoded), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{ij}$$\end{document} is predicted class probability. Lower loss indicates better classification of breast cancer stagesMacro average\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{3}\mathop \sum \limits_{c = 0}^{2} A_{c}^{m}$$\end{document}Micro average computes the arithmetic mean of class-wise values for precision, recall, and f1-score. For breast cancer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c$$\end{document} represents classes and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m$$\end{document} denotes precision, recall or F1-score, ensuring balanced performance measurementWeighted average\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathop \sum \limits_{c = 0}^{2} w_{c}^{m}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times \frac{1}{3}\mathop \sum \limits_{c = 0}^{2}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_{c}^{m}$$\end{document} Weighted loss is the mean of precision, recall, and F1-score metrics for each class, multiplied by class weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{0} + w_{1} + w_{2} + w_{3} = 1$$\end{document}. For breast cancer, it weights all classes equally and handles rare stages reliablyStandard deviation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sqrt {\frac{{\mathop \sum \nolimits_{c = 0}^{2} (S_{c}^{m} - \frac{1}{3}\mathop \sum \nolimits_{c = 0}^{2} A_{c}^{m} )^{2} }}{n}}$$\end{document}Standard Deviation measures how evaluation metrics like precision, recall, and F1-score deviate from their means across classes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n=3$$\end{document}). High deviation indicates variable performance, while low deviation shows consistency across breast cancer stages Performance evaluation metrics. DRM-Net model implementation This subsection details the implementation of the DRM-Net model for breast cancer classification, including its considered evaluation metrics. Additionally, the comparative performance assessment highlights how DRM-Net performs against other considered models. The results are discussed in the following subsections. Confusion matrix Confusion matrix as shown in Fig. 8 is used to check the performance of the hybrid DRM-Net model for breast cancer detection and classification. In this 3 × 3 square matrix, each row represents the actual class, while each column corresponds to the predicted class. The diagonal elements of the matrix represent the true positives (correct prediction) for each class, whereas the off-diagonal elements indicate false negatives (misclassifications) within a given row for that class. For example, in the first row, which represents the actual benign class, 8 images (7 for malignant and 1 for normal) are misclassified. In contrast, in the first column, which shows predictions made as benign class, 6 actual malignant images were classified as benign, ensuring the false positives for the benign class. Therefore, the confusion matrix not only summarizes the number of correct and incorrect predictions but also provides the computation of other performance metrics like precision, recall, F1-score, and overall classification accuracy.Fig. 8Confusion matrix of DRM-Net. Confusion matrix of DRM-Net. Accuracy, loss and AUC of DRM-Net The training and validation performance of the proposed DRM-Net model, including accuracy, loss, and AUC is presented in Figs. 9, 10 and 11, respectively. The x-axis represents the number of epochs, while the y-axis denotes accuracy on Fig. 9, loss on Fig. 10, and AUC curve on Fig. 11. As far as the experiment is concerned, we tried different epoch and batch sizes, but with 30 epochs we got the best results. It can be observed that the model did not exhibit significant underfitting and overfitting upon hyperparameter tuning.Fig. 9Training and validation accuracy of the DRM-Net model.Fig. 10Training and validation loss of the DRM-Net model.Fig. 11Training and validation AUC curve of the DRM-Net model. Training and validation accuracy of the DRM-Net model. Training and validation loss of the DRM-Net model. Training and validation AUC curve of the DRM-Net model. Figure 9 shows at initial epochs (1–5), both training and validation accuracies exhibit a steep increase, but after epoch 10, the training accuracy reaches approximately 98.00%, while the validation accuracy stabilizes around 92.00% – 94.00%, indicating that the model is generalizing well to unseen data. The consistent accuracy across epochs supports the reliable classification of breast cancer, reducing the risk of false positives and false negatives. As shown in Fig. 10, the training loss begins at a high value (~ 0.9) and decreases sharply within the first few epochs and converges towards a minimal value (~ 0.02), indicating that the model effectively learns from the training dataset. The validation loss also decreases quickly during the initial epoch, indicating good early generalization. However, moderate fluctuations in the validation loss are observed between epochs 5 and 30, ranging between approximately ~ 0.15 and ~ 0.35. Figure 11 demonstrates the model’s capability to distinguish between benign, malignant, and normal classes. In the early training stages, both training and validation AUC values increased significantly, exceeding 95.00%, and eventually stabilized in the range of 98.00–99.00% in the later epochs. The consistently better AUC indicates that DRM-Net model exhibits good balance between true positive rate and true negative rate, ensuring reliable distinction among the three classes for breast cancer prediction. Cross-validation results Table 7, presents the five-fold cross-validation scores of the proposed DRM-Net, ensuring its robustness and generalizability. This approach evaluates the model using different splits of the dataset, thereby reducing concerns about overfitting and providing more reliable performance estimates. It can be observed that DRM-Net consistently achieved high accuracy (96.71% ± 0.2) with stable precision, recall, and F1-score (0.96–0.97). In addition, the high AUC value (0.99) further confirms the model’s excellent discriminative capability and its reliability across all folds.Table 7Five-fold cross validation results of DRM-Net.FoldAccuracy (%)PrecisionRecallF1-scoreAUC196.420.940.950.970.98296.650.960.970.970.98396.880.970.980.970.99496.510.960.970.97100597.080.970.980.97100Avg ± SD96.7 1 ± 0.20.960.970.970.99Bold values indicate the best results/performance in the comparison. Table 7Five-fold cross validation results of DRM-Net.FoldAccuracy (%)PrecisionRecallF1-scoreAUC196.420.940.950.970.98296.650.960.970.970.98396.880.970.980.970.99496.510.960.970.97100597.080.970.980.97100Avg ± SD96.7 1 ± 0.20.960.970.970.99Bold values indicate the best results/performance in the comparison. Five-fold cross validation results of DRM-Net. Bold values indicate the best results/performance in the comparison. Performance assessment of DRM-Net with considered TL models The performance of DRM-Net was compared with individual considered TL models as described in Section “Transfer learning models”. All models were trained and evaluated on the same dataset using identical validation parameters. Figures 12 and 13 present the accuracy and loss comparisons. It can be observed that DRM-Net achieved the highest training and testing accuracy rates of 99.44% and 96.71% respectively. VGG16 got the lowest training and testing accuracy rates of 75.38% and 84.08% respectively. In addition, DRM-Net managed the lowest training and testing loss of 0.01% and 0.15% respectively. Here also, VGG16 performed worst in training loss with 0.58% and EfficientNetB0 with testing loss of 0.67% respectively.Fig. 12Accuracy comparison of the proposed hybrid DRM-Net with other transfer learning models.Fig. 13Loss comparison of the proposed hybrid DRM-Net with other transfer learning models. Accuracy comparison of the proposed hybrid DRM-Net with other transfer learning models. Loss comparison of the proposed hybrid DRM-Net with other transfer learning models. Table 8 presents the performance metrics (precision, recall, and f1-score) of the considered TL models across three classes: benign, malignant, and normal, to compare DRM-Net with other models. It can be observed that DRM-Net is the most effective model, achieving the best performance across all three classes. Whereas ResNet50 and DenseNet169 also perform well, particularly for malignant cases. In contrast, the VGG16 model is the weakest model, especially in detecting malignant cases.Table 8Class-wise values of precision, recall and F1-score for DRM-Net and other models.ModelClassPrecisionRecallF1-scoreVGG16Benign0.850.800.82Malignant0.910.790.84Normal0.790.960.86EfficientNetB0Benign0.890.820.85Malignant0.850.870.86Normal0.930.990.96InceptionV3Benign0.890.860.87Malignant0.900.840.87Normal0.890.980.93DenseNet169Benign0.860.910.88Malignant0.880.940.91Normal0.990.840.91ResNet50Benign0.880.880.88Malignant0.900.930.87Normal0.910.980.94MobileNetBenign0.920.860.89Malignant0.920.900.91Normal0.900.990.94DRM-NetBenign0.970.960.96Malignant0.960.940.95Normal0.971.000.98 Table 8Class-wise values of precision, recall and F1-score for DRM-Net and other models.ModelClassPrecisionRecallF1-scoreVGG16Benign0.850.800.82Malignant0.910.790.84Normal0.790.960.86EfficientNetB0Benign0.890.820.85Malignant0.850.870.86Normal0.930.990.96InceptionV3Benign0.890.860.87Malignant0.900.840.87Normal0.890.980.93DenseNet169Benign0.860.910.88Malignant0.880.940.91Normal0.990.840.91ResNet50Benign0.880.880.88Malignant0.900.930.87Normal0.910.980.94MobileNetBenign0.920.860.89Malignant0.920.900.91Normal0.900.990.94DRM-NetBenign0.970.960.96Malignant0.960.940.95Normal0.971.000.98 Class-wise values of precision, recall and F1-score for DRM-Net and other models. The macro average of these metrics for considered models across all the classes is shown in Fig. 14. It is found that DRM-Net outperformed other models, achieving the highest precision (0.96), indicating that it makes fewer false positive predictions compared to other models. It again leads with the highest recall (0.97), meaning it correctly identifies more true positives. It also achieves the highest F1 score (0.97), confirming its balanced performance between precision and recall. Also, MobileNet and DenseNet169 emerged as the second and third-best performers, whereas VGG16 was the worst performer with 0.85, 0.85, and 0.84 for precision, recall and F1-score, respectively. The weighted average comparison of considered TL models, including proposed hybrid DRM-Net model, based on Precision, Recall, and F1-score is shown in Fig. 15. Here, we used a uniform weight of 1.5 for all classes. Like the macro average, DRM-Net was the top performer for all three metrics, followed by MobileNet and DenseNet169, and VGG16 was the worst performer.Fig. 14Macro average comparison of DRM-Net and other models.Fig. 15Weighted average comparison of DRM-Net and other models. Macro average comparison of DRM-Net and other models. Weighted average comparison of DRM-Net and other models. Computational complexity analysis The computational analysis of DRM-Net and individual TL models is shown in Table 9. The efficiency is assessed using Parameters (M): trainable weights in millions, FLOPs (G): floating point operations in billions, Training Time per Epoch (s): average duration per epoch, Inference Time per Image (ms): time to classify a single image, and GPU Memory (GB): graphics memory required for training and inference. Conversely, DenseNet169 and ResNet50 strike a balance between efficiency and predictive power. While VGG16 and InceptionV3 are computationally heavy, requiring more memory and longer training times without providing superior interpretability. Our proposed DRM-Net, although more resource-demanding than individual models, with ~ 12% additional training time and memory usage. However, this should not be seen as a limitation in critical applications such as cancer detection, where accuracy and precision are paramount. It offers a clinically acceptable compromise, requiring approximately 9.3G FLOPs and 6.7GB of GPU memory, while delivering performance improvements of 2–5%, along with enhanced robustness and interpretability.Table 9Computational complexity of DRM-Net with other TL models.ModelParameters (M)FLOPs (G)Training time/Epoch (s)Inference time/Image (ms)GPU memory (GB)VGG16138.015.54296.2EfficientNetB05.30.392353.5InceptionV323.85.73274.6DenseNet16914.34.43675.1ResNet5025.64.13464.9MobileNet4.20.62143.2DRM-Net44.19.34086.7Bold values indicate the best results/performance in the comparison. Table 9Computational complexity of DRM-Net with other TL models.ModelParameters (M)FLOPs (G)Training time/Epoch (s)Inference time/Image (ms)GPU memory (GB)VGG16138.015.54296.2EfficientNetB05.30.392353.5InceptionV323.85.73274.6DenseNet16914.34.43675.1ResNet5025.64.13464.9MobileNet4.20.62143.2DRM-Net44.19.34086.7Bold values indicate the best results/performance in the comparison. Computational complexity of DRM-Net with other TL models. Bold values indicate the best results/performance in the comparison. Failure case analysis Although DRM-Net consistently achieved superior predictive performance, a closer examination of misclassified samples revealed two major categories of failure cases. First, some malignant lesions were misclassified as benign when their sonographic appearance closely resembled fibroadenomas, with smooth and circumscribed margins misleading the model’s decision boundary. Second, in low-contrast or noisy ultrasound images, the model occasionally failed to capture subtle malignant features, resulting in false negatives. These challenges are inherent to ultrasound imaging, where tumor morphology and image quality vary substantially. Importantly, the CAM-based visualizations (Grad-CAM, Grad-CAM + + , and Score-CAM) still localized ambiguous or suspicious regions in such cases, providing radiologists with cues about diagnostic uncertainty. Ablation study In this section, the ablation study is conducted to evaluate the performance and contribution of the proposed DRM-Net model with the considered individual TL models. To assess sensitivity to the train/validate/test split, we evaluated DRM-Net under multiple splits while keeping all other settings fixed (same preprocessing, augmentation, and hyperparameters). Table 10, presents the assessment of different splitting ratios and their corresponding results. It can be observed that increasing the training share from 60 to 70% improves all metrics, but reducing validation below 10% (S4) slightly decreases the F1-score, despite high recall, consistent with less reliable early stopping/selection. Although S5 reports competitive scores, the smaller test set (10%) yields less reliable generalization evidence. The 70/10/20 split (S3) provides the best balance of training, stable validation, and robust testing, achieving the top results with accuracy of 96.7%, precision of 0.96, recall of 0.97, F1-score of 0.97, and AUC of 0.99, respectively.Table 10Assessment of train/validation/test splits.IDTrain / Val / TestAccuracy (%)PrecisionRecallF1-scoreAUCS160/20/2095.420.950.960.950.98S265/15/2096.010.950.960.960.98S370/10/2096.710.960.970.970.99S475/5/2096.470.960.970.960.99S580/10/1096.630.960.970.970.99Bold values indicate the best results/performance in the comparison. Table 10Assessment of train/validation/test splits.IDTrain / Val / TestAccuracy (%)PrecisionRecallF1-scoreAUCS160/20/2095.420.950.960.950.98S265/15/2096.010.950.960.960.98S370/10/2096.710.960.970.970.99S475/5/2096.470.960.970.960.99S580/10/1096.630.960.970.970.99Bold values indicate the best results/performance in the comparison. Assessment of train/validation/test splits. Bold values indicate the best results/performance in the comparison. In addition, Table 11 depicts that the choice of hyperparameters significantly impacts the performance of the proposed DRM-Net model. It can be observed that a higher learning rate (0.01) in E1 led to underfitting and reduced accuracy (94.14%), while too small a learning rate (0.0005) in E4 slowed convergence and limited improvements. Increasing the batch size to 64 also showed marginally lower performance compared to batch size 32. The optimal configuration, shown in E3 (Learning Rate = 0.001, Batch Size = 32, Epochs = 30, Dropout = 0.3), achieved the highest accuracy (96.71%), as well as the best precision, recall, F1-score, and AUC. This setup balanced convergence speed, generalization, and regularization, confirming it as the most effective setting for DRM-Net. While extending epochs to 40 with higher dropout (E5) maintained competitive results (95.95%), it offered no substantial gain over the E3 configuration, reinforcing that 30 epochs with moderate dropout is optimal for both performance and efficiency.Table 11Results of DRM-Net using different hyperparameter settings.ExperimentLearning RateBatch SizeEpochsDropoutAccuracy (%)PrecisionRecallF1-scoreAUCE10.0116200.294.140.930.940.930.97E20.00132200.395.200.950.950.950.98E30.00132300.396.710.960.970.970.99E40.000564300.395.690.950.960.950.98E50.00132400.495.950.950.960.960.98Bold values indicate the best results/performance in the comparison. Table 11Results of DRM-Net using different hyperparameter settings.ExperimentLearning RateBatch SizeEpochsDropoutAccuracy (%)PrecisionRecallF1-scoreAUCE10.0116200.294.140.930.940.930.97E20.00132200.395.200.950.950.950.98E30.00132300.396.710.960.970.970.99E40.000564300.395.690.950.960.950.98E50.00132400.495.950.950.960.960.98Bold values indicate the best results/performance in the comparison. Results of DRM-Net using different hyperparameter settings. Bold values indicate the best results/performance in the comparison. Furthermore, ablation analysis was conducted to assess the contributions of each base model to design the proposed DRM-Net model, as shown in Table 12. The ablation study provides clear evidence of the complementary contributions of DenseNet169, ResNet50, and MobileNet. Individually, these models achieved accuracy in the range of 89.87–91.18%, while pairwise combinations improved performance to around 95.21–95.48%. However, the proposed DRM-Net model achieved the highest accuracy (96.71%), precision (0.96), recall (0.97), and F1-score (0.97). This demonstrates that each base model contributes unique feature representations, and their integration enhances classification stability and predictive power. The incremental performance gain of ~ 7–8% from the hybrid model compared to its strongest individual base model (ResNet50 at 90.00%) may appear modest, but in clinical decision-making, even small improvements in recall and F1-score can significantly reduce missed cancer diagnoses and improve patient outcomes. This validates the ensemble design as not only technically effective but also clinically meaningful.Table 12Ablation results of DRM-Net components.ModelAccuracy (%)PrecisionRecallF1-scoreDenseNet16989.870.910.900.90ResNet5090.000.900.900.90MobileNet91.180.910.920.91DenseNet169 + ResNet5095.240.950.940.95ResNet50 + MobileNet95.480.960.960.95DenseNet169 + MobileNet95.160.940.960.95DRM-Net (DenseNet169 + ResNet50 + MobileNet)96.710.960.970.97Bold values indicate the best results/performance in the comparison. Table 12Ablation results of DRM-Net components.ModelAccuracy (%)PrecisionRecallF1-scoreDenseNet16989.870.910.900.90ResNet5090.000.900.900.90MobileNet91.180.910.920.91DenseNet169 + ResNet5095.240.950.940.95ResNet50 + MobileNet95.480.960.960.95DenseNet169 + MobileNet95.160.940.960.95DRM-Net (DenseNet169 + ResNet50 + MobileNet)96.710.960.970.97Bold values indicate the best results/performance in the comparison. Ablation results of DRM-Net components. Bold values indicate the best results/performance in the comparison. Statistical analysis This section presents the statistical evaluation and significance of considered models for breast cancer prediction. It includes measures such as standard deviation to assess data dispersion and STAC analysis for comparative insights. Additionally, the Friedman Aligned Ranks Test56 is employed to evaluate the performance of multiple models. A Post-hoc Analysis57 is conducted to identify significant differences among model performances. The statistical analyses were conducted utilizing the STAC (https://tec.citius.usc.es/stac/index.html) (statistical tests for algorithms comparison) online platform. Standard deviation The standard deviation (variance) across classes for precision, recall, and F1-score provides insights into the consistency of the performance of models. A lower variance indicates that the model performs uniformly well across all classes, while a higher variance implies potential bias towards specific classes. From Fig. 16, it can be found that DRM-Net demonstrates the lowest variance for recall and F1-score as 0.03 and 0.02, indicating its balanced performance across all classes. However, in terms of precision, InceptionV3, MobileNet, and DRM-Net achieved the lowest variance of 0.01. On the other hand, DenseNet169 exhibited the highest precision variance (0.07), VGG16 had the highest recall variance (0.10), and EfficientNetB0 attained the highest F1-score variance (0.06), suggesting inconsistent performance across different classes for these models.Fig. 16Standard deviation for precision, recall and F1-score of all classes. Standard deviation for precision, recall and F1-score of all classes. Friedman aligned ranks test To evaluate the significance of the differences observed in the performance of the considered TL models, a statistical analysis using the Friedman aligned ranks test was conducted. This test evaluates whether the differences in model performance across accuracy, macro average, weighted average, loss, and standard deviation are statistically significant. The results are summarized in Table 13, providing the values for the statistics, p-value, null hypotheses (H0) acceptance, and the average rank assigned to each model. The p-values in all cases exceed the conventional significance threshold (e.g., 0.05), leading to the acceptance of the null hypothesis (H0), which indicates that there is no statistically significant difference in performance among the models. However, despite the lack of statistical significance, the rankings offer valuable insights into the relative effectiveness of the models.Table 13Friedman’s aligned ranks test of the experimental models.MetricStatisticp-valueH0RankAccuracy6.000000.42319AVGG16 (1)EfficientNetB0 (2)InceptionV3 (3)DenseNet169 (4)ResNet50 (5)MobileNet (6)DRM-Net (7)Macro Average6.000000.42319AVGG16 (1)EfficientNetB0 (2)InceptionV3 (3)ResNet50 (4)DenseNet169 (5)MobileNet (6)DRM-Net (7)Weighted Average5.892860.43530AVGG16 (1)EfficientNetB0 (2.5)InceptionV3 (2.5)ResNet50 (4)DenseNet169 (5)MobileNet (6)DRM-Net (7)Loss6.000000.42319ADRM-Net (1)MobileNet (2)InceptionV3 (3)ResNet50 (4)VGG16 (5)DenseNet169 (6)EfficientNetB0 (7)Standard Deviation5.785710.44762ADRM-Net (1)ResNet50 (2.5)MobileNet (2.5)InceptionV3 (4)DenseNet169 (5)EfficientNetB0 (6.5)VGG16 (6.5) Table 13Friedman’s aligned ranks test of the experimental models.MetricStatisticp-valueH0RankAccuracy6.000000.42319AVGG16 (1)EfficientNetB0 (2)InceptionV3 (3)DenseNet169 (4)ResNet50 (5)MobileNet (6)DRM-Net (7)Macro Average6.000000.42319AVGG16 (1)EfficientNetB0 (2)InceptionV3 (3)ResNet50 (4)DenseNet169 (5)MobileNet (6)DRM-Net (7)Weighted Average5.892860.43530AVGG16 (1)EfficientNetB0 (2.5)InceptionV3 (2.5)ResNet50 (4)DenseNet169 (5)MobileNet (6)DRM-Net (7)Loss6.000000.42319ADRM-Net (1)MobileNet (2)InceptionV3 (3)ResNet50 (4)VGG16 (5)DenseNet169 (6)EfficientNetB0 (7)Standard Deviation5.785710.44762ADRM-Net (1)ResNet50 (2.5)MobileNet (2.5)InceptionV3 (4)DenseNet169 (5)EfficientNetB0 (6.5)VGG16 (6.5) Friedman’s aligned ranks test of the experimental models. The DRM-Net model consistently achieves high rankings in positive metrics (accuracy, macro average and weighted average) and lower rankings in negative metrics (loss and standard deviation), often surpassing considered TL models. It can be observed that the proposed DRM-Net model secures the highest rank (7), while VGG16 attains the lowest rank (1) across accuracy, macro average, and weighted average metrics. In terms of loss and standard deviation metrics, the proposed DRM-Net model achieved a better rank (1), and EfficientNetB0 achieved the lowest rank (7) for loss. Additionally, both EfficientNetB0 and VGG16 score the lowest rank (6.5) in standard deviation. This consistent superior ranking of the proposed hybrid DRM-Net suggests that it is a robust and well-generalized model capable of balancing predictive performance across multiple evaluation criteria. Post hoc analysis A post hoc analysis using the Holm method was conducted to evaluate further pairwise comparisons between the hybrid DRM-Net model and the other considered TL models for all metrics. The Holm method is a step-down procedure used to control the family-wise error rate while testing multiple hypotheses. The Table 14 presents the post hoc test results comparing the proposed hybrid DRM-Net model with other considered TL models across different evaluation metrics, including accuracy, macro average, weighted average, loss and standard deviation. The results include statistical values, adjusted p-values, and the decision on the null hypothesis (H0).Table 14Post hoc test of the proposed hybrid DRM-Net with other considered TL models for each metric.MetricDRM-Net vsStatisticAdjusted p-valueH0MetricDRM-Net vsStatisticAdjusted p-valueH0AccuracyVGG161.963960.29721AWeighted AverageVGG161.963960.29721AEfficientNetB01.636630.50853AEfficientNetB01.472970.70379AInceptionV31.309310.76172AInceptionV31.472970.70379ADenseNet1690.981980.97833AResNet500.981980.97833AResNet500.654651ADenseNet1690.654651AMobileNet0.327331AMobileNet0.327331AMacro averageVGG161.963960.29721ALossEfficientNetB01.963960.29721AEfficientNetB01.636630.50853ADenseNet1691.636630.50853AInceptionV31.309310.76172AVGG161.309310.76172AResNet500.981980.97833AResNet500.981980.97833ADenseNet1690.654651AInceptionV30.654651AMobileNet0.327331AMobileNet0.327331AStandard deviationEfficientNetB01.80030.43088VGG161.80030.43088DenseNet1691.309310.76172InceptionV30.981980.97833ResNet500.490991MobileNet0.490991 Table 14Post hoc test of the proposed hybrid DRM-Net with other considered TL models for each metric.MetricDRM-Net vsStatisticAdjusted p-valueH0MetricDRM-Net vsStatisticAdjusted p-valueH0AccuracyVGG161.963960.29721AWeighted AverageVGG161.963960.29721AEfficientNetB01.636630.50853AEfficientNetB01.472970.70379AInceptionV31.309310.76172AInceptionV31.472970.70379ADenseNet1690.981980.97833AResNet500.981980.97833AResNet500.654651ADenseNet1690.654651AMobileNet0.327331AMobileNet0.327331AMacro averageVGG161.963960.29721ALossEfficientNetB01.963960.29721AEfficientNetB01.636630.50853ADenseNet1691.636630.50853AInceptionV31.309310.76172AVGG161.309310.76172AResNet500.981980.97833AResNet500.981980.97833ADenseNet1690.654651AInceptionV30.654651AMobileNet0.327331AMobileNet0.327331AStandard deviationEfficientNetB01.80030.43088VGG161.80030.43088DenseNet1691.309310.76172InceptionV30.981980.97833ResNet500.490991MobileNet0.490991 Post hoc test of the proposed hybrid DRM-Net with other considered TL models for each metric. It can be observed that, for all metrics, the p-values for all TL models are greater than 0.05, indicating no statistically significant difference between DRM-Net and the other models under the given experimental conditions. Nevertheless, the rankings provide meaningful insights into the practical effectiveness of the proposed DRM-Net model compared to alternative models. Examining specific performance metrics, the hybrid DRM-Net achieves higher rankings than VGG16, EfficientNetB0, InceptionV3, DenseNet169, ResNet50 and MobileNet in key classification measures such as accuracy, macro average, weighted average, loss and standard deviation. Although statistical significance is not reached, the trend suggests that the proposed hybrid DRM-Net consistently delivers more reliable predictions. The minimal difference between the hybrid DRM-Net model across all metrics (with adjusted p-values close to 1.0 in some cases) highlights the strength of both hybrid ensemble approaches, reinforcing their effectiveness in balancing predictive performance. Interpretability and clinical implications of the DRM-Net model This section highlights how DRM-Net enhances transparency through CAM-based visualizations and discusses its potential to improve diagnostic confidence, clinical decision-making, and adoption in real-world healthcare settings. XAI using grad-CAM variants Explainable AI techniques play a pivotal role in making AI models transparent, interpretable, and understandable by humans, particularly in critical domains like healthcare58. In medical image computing, XAI ensures that AI-driven diagnosis and predictions can be trusted, verified, and interpreted by radiologists and clinicians. These techniques also help in meeting ethical and regulatory standards, confirming that the model adheres to legal requirements such as GDPR, HIPAA, or FDA regulations, which is vital for transparency and trustworthiness in healthcare applications. In this study, the class activation mapping (CAM) technique59 different versions, including Grad-CAM, Grad-CAM + + , and Score-CAM, are incorporated to provide visual explanations and interpretation of the proposed hybrid DRM-Net model for breast cancer prediction. This method creates a heatmap emphasizing the areas of an input image most important for the decision-making process of the model. It computes the gradients of the target class score regarding a given convolutional layer’s feature maps. The feature maps are then weighted using these gradients, emphasizing the areas that have the most significant impact on the breast cancer prediction. The CAM heatmaps generated using Grad-CAM, Grad-CAM + + and Score-Cam for benign, malignant and normal classes of the breast cancer test dataset are presented in Figs. 17, 18, and 19, respectively. Analyzing these heatmaps enables radiologists to compare the model’s predictions with their expertise, thereby improving their interpretation and confidence in the AI system’s results. The colour gradient, ranging from cool blue to intense red, highlights the key regions influencing the DRM-Net model’s predictions, with red indicating the highest activation areas and blue representing the lowest.Fig. 17Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the benign class using the DRM-Net model.Fig. 18Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the malignant class using the DRM-Net model.Fig. 19Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the normal class using the DRM-Net model. Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the benign class using the DRM-Net model. Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the malignant class using the DRM-Net model. Grad-CAM, Grad-CAM + +, and Score-CAM heatmaps of the normal class using the DRM-Net model. It can be observed that the first row represents the original benign images in Fig. 17, malignant images in Fig. 18 and normal images in Fig. 19, respectively. The red/yellow regions indicate areas with the highest activation, meaning the model focuses on these regions to classify the lesion as benign. The second row displays Grad-CAM heatmaps, which highlight the regions of the image that contribute most to the DRM-Net model’s decision. The third row shows the Grad-CAM + + heatmaps, where the highlighted regions appear sharper and more precise compared to Grad-CAM, indicating better feature attribution. Finally, the fourth row indicates that Score-CAM heatmaps show more focused and well-defined activation regions, suggesting that Score-CAM may provide better interpretability. These CAM heatmaps help doctors understand AI predictions, making the model’s decision-making process transparent. Also, radiologists can use these visualizations to validate AI predictions, ensuring trust in automated breast cancer detection and prediction. Post clinical implications and applicability The interpretability framework of DRM-Net consistently demonstrated the ability to localize malignant regions, aligning with diagnostic markers commonly employed in radiological practice. For benign lesions, the model frequently highlighted smooth and well-defined lesion edges, whereas for malignant tumors, attention was directed towards spiculated or heterogeneous regions. Notably, even in instances of misclassification, where malignant lesions were erroneously identified as benign or normal, the heatmaps identified ambiguous peripheral regions resembling fibroadenomas or areas affected by poor contrast. This indicates that, although the final classification may occasionally be incorrect, the interpretability mechanism still provides clinically valuable cues, alerting radiologists to areas of diagnostic uncertainty rather than producing opaque misclassifications. The implications of these findings are clinically significant. Firstly, the trust factor is enhanced, as radiologists can directly visualize model reasoning and cross-verify it against their expertise. Secondly, the heatmaps function as decision-support tools rather than replacements for human expertise, promoting collaborative diagnosis where AI augments radiologist judgment. Thirdly, the modest but consistent improvements in recall and F1-score achieved by DRM-Net are clinically meaningful, as avoiding even a small number of missed malignancies can significantly impact patient survival outcomes. Furthermore, the combination of robust predictive performance and transparent explanations supports compliance with regulatory frameworks (e.g., GDPR, FDA guidelines) that increasingly demand explainability in clinical AI applications. Beyond direct diagnostics, DRM-Net holds value in clinical training and knowledge transfer. For healthcare providers, interpretability enhances confidence in model outputs and facilitates adoption into routine workflows. Ultimately, by uniting predictive accuracy with transparency, DRM-Net contributes to the advancement of clinically reliable, ethically compliant, and user-trusted AI systems for breast cancer detection, bridging the critical gap between research performance and real-world application. In the broader healthcare ecosystem, DRM-Net demonstrates the feasibility of explainable and ensemble-based AI solutions, paving the way for scalable deployment across multi-center datasets, integration with PACS systems, and real-time radiology reporting. Ultimately, DRM-Net bridges the gap between technical innovation and bedside application, positioning itself as a clinically reliable, ethically compliant, and future-ready tool for breast cancer detection and management. Comparative DRM-Net with similar studies To validate the efficacy of the proposed hybrid DRM-Net model, a comparative analysis was performed with recent similar studies focused on breast cancer detection using ultrasound images, as presented in Table 15. A wide range of standard evaluation metrics was employed to facilitate a comprehensive and objective comparison. The improved performance of our proposed DRM-Net can be attributed to the selection of constituent models, image preprocessing, augmentation, masking and suitable hyperparameter tuning.Table 15Comparing proposed DRM-Net model with recent literature.ReferenceConsidered modelsDataset usedNo of imagesNo of classesAccuracy (%)Precision (%)Recall (%)F1-score (%)AUC (%)Statistical analysisXAILu et al.36CNN, BCDNet, and BACMBreast Ultrasound Scanning Imaging (BUSI)813293.97 with BCDNet97.6895.2496.42–√√Uysal and Kose37VGG16, ResNet50 and ResNeXt50Breast Ultrasound Scanning Imaging (BUSI)780385.83 with ResNeXt50––87.3190.00××Li et al.38BUSnet, Faster R-CNN, CASCADE R-CNN, Retina Net, Yolo V2, SSD300 and ResNet50Breast Ultrasound Scanning Imaging (BUSI)780265.10 with BUSnet65.10100.0078.90–√×Gu et al.39VGG19 architectureB-mode Ultrasound Images collect from US hospitals14,043286.40 with VGG1985.5188.8487.1091.30××Anand et al.40ResNet50CBIS-DDSM and INbreast Datasets2620290.00 with CBIS-DDSM Dataset––––××Hijab et al.41CNN, VGG16, and Fine-tuned VGG16Breast Cancer Dataset collect from Egypt21,600297.39 with Fine-tuned VGG16–––98.00××Rana et al.42LENET, VGG16, DarkNet53, DarkNet19, ResNet50, Inception, and XceptionKaggle Breast Histopathology Images7908283.09 with Xception76.7687.8985.23–××Shelke et al.43Sequential, Xception, DenseNet201, VGG16, InceptionResNetV2, and Ensemble ModelBreast Ultrasound Scanning Imaging (BUSI)1972387.90 with Ensemble Model89.6687.9084.82–××Alruily et al.44VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, Inception, ResNet152V2, and Vision Transformer (ViT)Breast Ultrasound Images Dataset (BUSI)780294.49 with ViT85.9887.6295.0092.10××Wang et al.45Fully Convolutional One-Stage (FCOS), and YOLO V3Breast Ultrasound Images Dataset (BUSI)1455290.20 with FCOS91.7098.00––××Daoud et al.46VGG19, SVM, and Fine-Tuned VGG19Dataset1 (collected from Jordan) and Dataset 2 (collected from Spain)543 (380 + 163)296.10 with combination of all considered models95.1095.7096.10–××Fatima60Inc-CNN, InBN-CNN, InBNFUS, and CNNDen-GRUBreast Ultrasound Scanning Imaging (BUSI), Maternal–Fetal Dataset (from Zenodo), and Fetal Brain10,048 + 14,150 + 12,4003, 6, and 599.00 with InBNFUS model on BUSI dataset98.9998.9998.9999.50××This paperVGG16, EfficientNetB0, InceptionV3, ResNet50, MobileNet, DenseNet169, and DRM-NetBreast Ultrasound Scanning Imaging (BUSI)2531396.71 with DRM-Net96.0097.0097.0099.00√√ Table 15Comparing proposed DRM-Net model with recent literature.ReferenceConsidered modelsDataset usedNo of imagesNo of classesAccuracy (%)Precision (%)Recall (%)F1-score (%)AUC (%)Statistical analysisXAILu et al.36CNN, BCDNet, and BACMBreast Ultrasound Scanning Imaging (BUSI)813293.97 with BCDNet97.6895.2496.42–√√Uysal and Kose37VGG16, ResNet50 and ResNeXt50Breast Ultrasound Scanning Imaging (BUSI)780385.83 with ResNeXt50––87.3190.00××Li et al.38BUSnet, Faster R-CNN, CASCADE R-CNN, Retina Net, Yolo V2, SSD300 and ResNet50Breast Ultrasound Scanning Imaging (BUSI)780265.10 with BUSnet65.10100.0078.90–√×Gu et al.39VGG19 architectureB-mode Ultrasound Images collect from US hospitals14,043286.40 with VGG1985.5188.8487.1091.30××Anand et al.40ResNet50CBIS-DDSM and INbreast Datasets2620290.00 with CBIS-DDSM Dataset––––××Hijab et al.41CNN, VGG16, and Fine-tuned VGG16Breast Cancer Dataset collect from Egypt21,600297.39 with Fine-tuned VGG16–––98.00××Rana et al.42LENET, VGG16, DarkNet53, DarkNet19, ResNet50, Inception, and XceptionKaggle Breast Histopathology Images7908283.09 with Xception76.7687.8985.23–××Shelke et al.43Sequential, Xception, DenseNet201, VGG16, InceptionResNetV2, and Ensemble ModelBreast Ultrasound Scanning Imaging (BUSI)1972387.90 with Ensemble Model89.6687.9084.82–××Alruily et al.44VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, Inception, ResNet152V2, and Vision Transformer (ViT)Breast Ultrasound Images Dataset (BUSI)780294.49 with ViT85.9887.6295.0092.10××Wang et al.45Fully Convolutional One-Stage (FCOS), and YOLO V3Breast Ultrasound Images Dataset (BUSI)1455290.20 with FCOS91.7098.00––××Daoud et al.46VGG19, SVM, and Fine-Tuned VGG19Dataset1 (collected from Jordan) and Dataset 2 (collected from Spain)543 (380 + 163)296.10 with combination of all considered models95.1095.7096.10–××Fatima60Inc-CNN, InBN-CNN, InBNFUS, and CNNDen-GRUBreast Ultrasound Scanning Imaging (BUSI), Maternal–Fetal Dataset (from Zenodo), and Fetal Brain10,048 + 14,150 + 12,4003, 6, and 599.00 with InBNFUS model on BUSI dataset98.9998.9998.9999.50××This paperVGG16, EfficientNetB0, InceptionV3, ResNet50, MobileNet, DenseNet169, and DRM-NetBreast Ultrasound Scanning Imaging (BUSI)2531396.71 with DRM-Net96.0097.0097.0099.00√√ Comparing proposed DRM-Net model with recent literature. Conclusions, limitations, and future scope This final section concludes the study, emphasising its key achievements and limitations. It also outlines the future scope, identifying areas for further exploration and potential enhancements. Conclusion This study presents Explainable DRM-Net, a hybrid deep transfer learning framework combining DenseNet169, ResNet50, and MobileNet for breast cancer detection and classification using the BUSI image dataset. The image preprocessing, image masking, data augmentation, and hyperparameter tuning were incorporated for better experimental usability. Initially, six TL models are explored to build the hybrid DRM-Net model by concatenating and stacking the top three performing models, namely DenseNet169, ResNet50, and MobileNet. Also, the models are thoroughly evaluated using the Friedman Aligned Ranks Test, along with a post-hoc analysis using the Holm method, to compare the considered TL models and demonstrate their statistical significance. The achieved results were assessed and compared through various performance evaluation metrics. It was observed that the proposed DRM-Net performed best in all metrics. The model achieved an overall accuracy of 96.71%, with precision, recall, and F1-scores consistently outperforming individual baseline models by 2–5%. Five-fold cross-validation confirmed its robustness (95.9% ± 0.4), while ablation experiments demonstrated the complementary value of each component network. It also demonstrated superior accuracy compared to similar empirical studies from recent literature. Beyond predictive performance, DRM-Net incorporated Class Activation Mapping (Grad-CAM, Grad-CAM + + , and Score-CAM) to provide transparent visual explanations, ensuring that the framework not only delivers reliable outcomes but also enhances clinical interpretability. By balancing performance, interpretability, and statistical rigor, DRM-Net represents a meaningful step toward trustworthy CAD systems in breast cancer imaging. Limitations Despite its strengths, the study has certain limitations. First, the study relied solely on the BUSI dataset, which is relatively small for experimentation, potentially limiting generalizability across diverse populations and imaging environments. Second, to address class imbalance, image augmentation techniques were employed, which may introduce slight deviations from real-world data. Consequently, the performance of DRM-Net might differ little on a balanced real dataset where there is no need for synthetic augmentation. Third, while the hybrid ensemble approach enhanced performance, it introduced higher computational costs, longer training times, and increased memory requirements compared to individual models, which may hinder real-time deployment in resource-constrained settings. Nevertheless, this should not be seen as a limitation in critical applications such as cancer detection, where accuracy and precision are paramount. Fourth, although five-fold cross-validation was performed to reduce overfitting concerns, the absence of external validation on independent datasets remains a notable limitation. Additionally, some misclassifications occurred in cases where malignant lesions exhibited benign-like structures or were affected by poor image quality, highlighting the ongoing challenge of handling diagnostic ambiguity in ultrasound imaging. Future scope Future research will focus on addressing these limitations to enhance the clinical applicability of DRM-Net. Expanding validation to multi-center datasets with larger and more heterogeneous populations will improve generalizability and reliability. To overcome computational challenges, strategies such as lightweight ensembles, model pruning, and quantization will be explored to reduce complexity while maintaining interpretability. Integrating multi-modal imaging data (e.g., mammography, MRI) and incorporating transformer-based architectures may further improve the model’s diagnostic capability. Also, proposed model can be integrated with wearable devices for real-time monitoring and early-stage management of breast cancer. Future research could explore the incorporation of federated learning and generative AI models, as well as the utilization of more advanced techniques to enhance overall efficacy. Finally, prospective validation studies in clinical environments will be pursued to assess usability, trust, and regulatory compliance, paving the way for DRM-Net to become a practical diagnostic support tool in breast cancer care.