A multimodal deep learning framework for precise silicosis detection on radiographic images Authors: Ramesh N S V S C Sripada, Pothuri Surendra Varma, E Laxmi Lydia, K Radhika, Jang Mook Kang, Gyanendra Prasad Joshi, Cheolhee Yoon Published: 2026 ABSTRACT AbstractPneumoconiosis, a common occupational lung illness, arises from inhaling dust, with silicosis specifically caused by fine crystalline silica dust, leading to lung scarring and inflammation. The diagnosis of silicosis relies on routine monitoring, which includes physical examinations, medical history reviews, and imaging. Chest radiography is a common form of medical screening due to its affordability, efficiency, and suitability for routine use. Recent success in deep learning (DL) for medical image classification has demonstrated that DL algorithms can identify silicosis with high precision by classifying CT images. DL models, specifically convolutional neural networks, have become an effective approach due to their ability to analyse medical images. Therefore, this study presents a novel Multimodal Deep Learning Framework for Early Identification of Silicosis Diagnosis (MDLF-EISD) using radiological images, focused on enabling timely clinical intervention and improving patient outcomes. The proposed framework applies feature fusion (EfficientNet-B3, a capsule network, and ConvNext V2) for integrating complementary radiographic representation, enhancing the ability of the model in capturing disease-specific patterns over different levels of silicosis. Moreover, a convolutional bidirectional attention model is utilised to classify silicosis into corresponding categories effectively. An extensive simulation studies were carried out to evaluate the enhanced performance of the MDLF-EISD method under Silicodata. The comparative analysis of the MDLF-EISD method illustrated a superior accuracy value of 98.73% over other models. These results demonstrate that the feature fusion improves discriminative capabilities for silicosis-related radiographic findings. The proposed system has the ability to support as a computer-aided screening tool for early silicosis diagnosis, particularly in resource-limited clinical and occupational health settings where access to expert radiologists is limited.Keywords: Precision diagnostics, Radiological image analysis, Silicosis diagnosis, Feature fusion, Artificial intelligence, Deep learningSubject terms: Computational biology and bioinformatics, Diseases, Engineering, Health care, Mathematics and computing, Medical research Introduction Silicosis is described as a debilitating and chronic pneumoconiosis, etiologically categorised as a wide range of interstitial lung disease (ILD). Its pathogenesis is connected to a chronic inhalation of silica dust, which leads to a gradual pulmonary fibrosis, irreparable respiratory failure, and demise1. Traditionally related to higher-risk environments such as quarrying and mining, silicosis has been established as a significant challenge to people’s health. However, in the last few years, silicosis has undergone a worldwide resurgence, which marks a changeover in the primary source of revelation2. As the primary pulmonary indication, simple silicosis (SS) is represented by the diffusion of tiny, distinct fibrotic nodules in parenchymal tissues3. This way can lead rapidly to the elevation to progressive massive fibrosis (PMF), described by the permanent coalescence of fibrotic lesions into huge conglomerate groups, therefore influencing severe and permanent damage to lung functions4. The latency period is short in ES-related silicosis when compared with traditional silicosis. The primary procedures for analysing silicosis are clinical and occupational history, chest X-rays (CXRs), higher-resolution computed tomography (HRCT), and pulmonary function tests (PFTs)5. In spite of current techniques accessible in medical applications, the early and precise recognition of silicosis is one of the significant challenges6. Intra- and inter-observer variation in the analysis of CXRs, the complication in understanding radiological results like tiny opacities, in the earlier disorder stages, restricted access to clinical image tests, higher cost-effectiveness, and the HRCT radiation vulnerability make it necessary to determine novel cost-effective tools that support analysis7. Moreover, a few recent studies demonstrate that assured biomarkers extracted from routine blood tests, like angiotensin-converting enzyme (ACE), fibrinogen, lactate dehydrogenase (LDH), and several cellular ratios, are associated with the chronic inflammatory condition and might help in differentiating patients with SS from healthy persons3. These biomarkers may serve as an extra element to an analysis performed through imaging methods, even though the possible role of these biomarkers is currently not recognised8. The fast evolution of artificial intelligence (AI) has led to wide applications of deep learning (DL) methods for the analysis of medical images. The methods mentioned above use a multi-layer network to assist the automatic learning of implied relations within the data9. The resultant features are frequently different and significant, especially in the application of silicosis images. DL can offer unsupervised or semi-supervised independent learning of a target image for a classification task10. Also, it can analyse images with similar features and emulate the autonomous learning and analysis abilities of human beings, therefore decreasing the individuality of feature extraction11. Currently, no appropriate reports on the valuation of pneumoconiosis image analysis depend on the DL computer-aided diagnostic (CAD) system technique. Deep residual networks (DRNs) and Convolutional neural networks (CNNs) are standard DL methods utilised for the classification of images. Such techniques give the benefits of generalisability, practicability, and simplicity12. Early-stage silicosis often presents with subtle opacities that overlap with tuberculosis and STB, leading to diagnostic ambiguity in CXR. Existing CNN-based CAD systems struggle to retain spatial hierarchies and global contextual features needed to distinguish these conditions. Therefore, integrating EfficientNet (local features), CapsNet (pose-aware spatial relations), and ConvNeXt V2 (global response normalisation) provides a more biologically meaningful representation. Thus, the multimodal DL model improves precision diagnostics by enabling finer discrimination between silicosis, TB, STB, and normal CXRs, overcoming common clinical interpretive challenges. To accurately identify and classify silicosis, a kind of pulmonary fibrosis, a lung condition triggered by breathing in tiny bits of silica, this paper proposes a novel Multimodal Deep Learning Framework for Early Identification of Silicosis Diagnosis (MDLF-EISD) by utilising radiological images. The proposed model aims to achieve early and precise identification and classification of silicosis, facilitating appropriate clinical intervention and increasing patient outcomes. This study follows a multi-faceted approach encompassing a refined image pre-processing pipeline through resizing, normalisation, gamma correction, and noise removal, multimodal feature fusion, and classification. The proposed MDLF-EISD framework is evaluated on benchmark datasets using several performance metrics. This study hypothesises that combining heterogeneous feature extractors with a bidirectional attention classifier can improve early-stage silicosis detection in chest radiographs. The main contributions are: Design a novel multimodal feature fusion approach integrating three deep learning architectures: EfficientNet-B3, capsule network (CapsNet), and ConvNext V2, achieving efficient and robust representation learning. Design a novel multimodal feature fusion approach integrating three deep learning architectures: EfficientNet-B3, capsule network (CapsNet), and ConvNext V2, achieving efficient and robust representation learning. Integrate a convolutional bidirectional attention, which is the integration of a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and attention mechanism (AM), capturing effective dependencies and performing silicosis classification tasks. Integrate a convolutional bidirectional attention, which is the integration of a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and attention mechanism (AM), capturing effective dependencies and performing silicosis classification tasks. Conduct extensive simulations and experiments and exhibit the superiority of the presented method on the Silicodata benchmark, a public dataset for silicosis detection, across diverse performance measures. Conduct extensive simulations and experiments and exhibit the superiority of the presented method on the Silicodata benchmark, a public dataset for silicosis detection, across diverse performance measures. The structure of the manuscript is as follows. Section "Review of current diagnostic approaches for silicosis" includes a literature review concerning the techniques utilised to detect and classify silicosis using radiological images. Section "Methodological framework" provides the methodology through the presentation of the dataset used, feature representation strategy, classification approaches, and the metrics utilised to validate the model performance. Section "Results interpretation" outlines the experimental procedure and corresponding results. This section includes a comprehensive investigation of the model’s performance across numerous measures. Section "Conclusion" concludes the article by summarising its key contribution. Review of current diagnostic approaches for silicosis This section offers a concise review of the related works regarding the diagnosis of silicosis using radiological images. Bui et al.13 introduced extensive research on the classification and identification of silicosis-based lung diseases. Initially, the development of a recently constructed CXR image data called SVBCX, which is designed for the features of lung diseases induced by different factors, offering a valuable resource for the silicosis and pneumonia scientific community, and after developing an innovative DL method that incorporates graph transformer models with a conventional DNN architecture for effectively classifying pneumonia and silicosis. Sharma et al.14 developed a technique for detecting silicosis with the help of transfer learning methods under X-ray radiographs. Moreover, the Grad-CAM method creates heatmap images that emphasise the significant extracted features from the CXR. This technique can work as a possible alternative with reduced accessibility to particularised radiological equipment and experts’ radiology specialists for identifying silicosis on CXR. The identification of silicosis on CXR greatly depends on the radiology professionals, which leads to slow diagnosis and detection. Xuan et al.15 intended to evaluate the effectiveness of a nanofiber sensor array as well as pattern identification for noninvasively and rapidly identifying silicosis. Silicosis issues constitute various breathing patterns from healthy controls, classification outcomes by applying a higher constant with the diagnosis of medical experts. Breath assessments worked with the sensor array and pattern recognition have been predicted for offering a fast, consistent recognition for silicosis. This technique provides various types of features, diverse methods, and data sets over extensive time periods that can be referred to for the silicosis expiratory analysis method. Liu et al.16 intended to monitor the detection of biomarkers of silicosis depending on the distinguishingly indicated serum proteins among silicosis fibrosis-affected individuals and normal healthy people. The patients with medially diagnosed silicosis have been screened, which has developed into the basis of treatment classes. Likewise, 20 healthy patients aged 20 years visiting to check-up for physical analysis have been chosen. Serum samples are acquired, and an integrated protein chip with surface-enhanced laser desorption ionisation flight mass spectroscopy is implemented for performing the analysis of serum. Akgün et al.17 implemented a study to assess how the ILO classification was utilised, but acquired CXR for screening with the accomplishment of claim data by the inquiry points, acquired from the ILO reading. The authors employed an experienced reference panel containing 5 ILO readers to re-analyse each and every CXR individually through ILO standard images and gained an outcome derived from the understandings between at least three readers. Chen et al.18 presented a PFD delivery model (PFD-Exo), macrophage-derived exosomes. Initially, PFD can be provided by sonication, and after, PFD-Exo is represented through UV absorption and Raman spectral imaging spectroscopy. Ultimately, in vitro and vivo silicosis techniques have been formed to analyse their antifibrotic impacts. During a mouse model of silicosis, PFD-Exo has been determined to be collected in the lungs and after intratracheal administration, resulting in majorly increased pulmonary disorders and fibrosis, with decreased adverse effects on the gastrointestinal tract. Akhter et al.19 introduced the SHIELD - a self-supervised, silicosis-centred hierarchical imaging model for earlier diagnosis of lung diseases. This technique exploits a multi-resolution jigsaw puzzle pretext task under CXR imagery for extracting and maintaining attributes in the evaluation of lung regions. With the help of utilising a pyramidal approach to design pretrained models on different resolutions, after that, fine-tuning and a two-stage ensembling through various DL methods, SHIELD reaches improved diagnostic precision. The above-mentioned progress analysis of medical images facilitates UN Sustainable growth. Xiong and Tang20 proposed a synthetic image generation technique that combines virtual settings of form dust to create training data. Recently, three object detection methods, comprising single-shot detection (SSD), you only look once (YOLO), and Faster-RCNN, have been trained by employing only synthetic images. Ultimately, this model offers a comparison evaluation of object identification techniques for real-time dust observation with respect to computational effectiveness and accuracy. A summary of previous studies on existing models is presented in Table 1. Table 1A tabular representation of silicosis diagnosis models.ReferencesTechniquesOutcomesAdvantagesDisadvantagesBui et al.13Ensemble learningAUC of 0.99.A robust hybrid model that achieves highly accurate lung-inflammation classification.The approach is computationally complex and requires substantial resources.Sharma et al.14Deep transfer learning models,F1 score of 0.8166.Offers an easier and more accessible way to detect silicosis from X-rays with better accuracy.A limited number of X-ray images reduces reliability across different cases.Xuan et al.15XGBoost and DNN models.Accuracy of 96%.Provides a fast, noninvasive, and highly accurate method for detecting silicosis through breath analysis.Relies on specialised sensor equipment that may not be widely accessible.Liu et al.16Principal Component AnalysisAchieved better performance.Identifies specific serum protein differences, helping in early detection and understanding of silicosis.Small sample size limits the reliability and broader applicability of the findings.Akgün et al.17DL approachAttained superior outcomesAn expert re-evaluation of silicosis cases, highlighting discrepancies and potential improvements in the ILO classification process.The ILO classification can be subjective and inconsistently applied, leading to possible misdiagnosis and compensation issues.Chen et al.18NAGained enhanced resultsThe model improves drug delivery, effectively reduces lung fibrosis, and minimises side effects.Requires specialised exosome preparation, making it complex and difficult to implement clinically.Akhter et al.19DL modelsAccuracy of 72%.The method enables cost-effective early detection of silicosis using limited data.May still miss some cases, limiting reliability for clinical use.Xiong and Tang20DL methodsPrecision of 97.63%.Uses synthetic images to train models, enabling effective dust detection even with limited real-world data.Performance may be lower in highly complex or unpredictable real-world construction environments. A tabular representation of silicosis diagnosis models. Present studies on the diagnosis of silicosis employing radiological images mainly based on chest X-ray (CXR) data and frequently concentrates on limited classification challenges, like distinctive silicosis from normal cases. But, in real medical settings, silicosis should be distinguished from other pulmonary circumstances like silico-tuberculosis and tuberculosis that display similar radiographic features. Furthermore, the availability of huge and well-balanced datasets for silicosis remains limited that affects the normalization and robustness capability of DL models. Several existing methods also function as black-box systems, presenting high accuracy but lacking necessary interpretability for clinical decision-making. Consequently, there is a requirement for more dependable and understandable automated systems, which can efficiently analyze radiological images and support precise multi-class diagnosis of silicosis and relevant lung diseases. The existing silicosis detection methods have illustrated promising performance, many challenges remain. Ensemble and DL techniques have attained high classification accuracy; Moreover, many researches struggle with computational complexity, limited dataset sizes, dependence on specialised equipment, and reduced generalisability over various clinical settings. Moreover, stand-alone model may not efficiently acquire the difficult patterns related with silicosis and related pulmonary diseases. Several techniques also show limited reliability for clinical decision support because of the insufficient feature representation and inconsistency in image properties. Motivated by these limitations, this article presents the MDLF-EISD architecture to improves the proficiency and effectiveness of silicosis detection from chest X-ray images. Methodological framework The proposed model follows a systematic approach to identify silicosis using radiological imagery that blends the process of image pre-processing, multimodal feature extraction, classification, and performance analysis. A benchmark (Silicodata: An Annotated Benchmark CXR Dataset for Silicosis Detection). Figure 1 depicts the overall structural design of the MDLF-EISD framework. As demonstrated in Fig. 1, the proposed MDLF-EISD method proceeds through a sequence of phases. Overall structural design of the MDLF-EISD model. Pre-processing In the initial stage, pre-processing is performed, involving image resizing, pixel intensity normalisation, gamma correction, and noise removal. These steps ensure that the inputs are appropriate as well as ready for the subsequent feature extraction stage. Feature extraction In this stage, a fusion of three advanced CNN architectures is utilised. EfficientNet-B3: Extracting rich and hierarchical features from images. EfficientNet-B3: Extracting rich and hierarchical features from images. CapsNet: It focuses on preserving spatial hierarchies and relationships between features. CapsNet: It focuses on preserving spatial hierarchies and relationships between features. ConvNext V2: Captures deep and discriminative features through its modernised convolutional design. ConvNext V2: Captures deep and discriminative features through its modernised convolutional design. Classification A convolutional bidirectional attention module is designed to perform accurate detection and classification of silicosis, where CNN extracts spatial features, the BiLSTM captures temporal dependencies, and the AM enriches the model’s focus on the most relevant features, ultimately leading to robust classification of silicosis based on the extracted highly discriminative attributes. Model validation Performance assessment of the proposed system is conducted on a benchmark silicosis detection dataset. Comparative results pointed out the superiority of the model across recent approaches. Dataset used The experimental analysis of the proposed model is examined under Silicodata21. This study offers a public dataset for silicosis recognition. The dataset consists of frontal CXR gathered over three years. The data are collected from stone workers in the basic health centres of Rajasthan, India. This dataset holds 3044 instances in total under four classes, like silicosis, STB, Tuberculosis, and normal, as presented in Table 2. Figure 2 depicts the sample images. The Silicodata CXR images have an average resolution of approximately 3567 × 2898 pixels, and the dataset contains variable image sizes because of the different acquisition sources, which were resized during the pre-processing stage for model training. Table 2Details of the dataset.ClassesTotalSilicosis1057STB752Tuberculosis750Normal481Total3044 Details of the dataset. Sample images. Pre-processing workflow Image pre-processing is a crucial step as it prepares the inputs by enhancing their quality for subsequent analysis. In this study, comprehensive pre-processing steps are followed, namely image resizing, pixel intensity normalisation, gamma correction, and denoising to enhance image clarity and quality. Image resizing It is a vital step in image pre-processing for CV and ML applications. Its primary objective is to ensure that every input image shares a uniform size to match the fixed input dimensions essential for approaches. Image resizing can regularise the data and enable batch processing and effective model training. Normalisation Image normalisation has been employed for upgrading the outcome, speed of convergence, and steadiness of the method by modifying the distribution and range of pixel values in the image22. The utilisation of diverse approaches of image normalisation is proven to impact the execution of ML. There are three commonly utilised approaches of image normalisation, comprising empirical model, z-score and min-max normalisation. The min-max normalisation approach can process every pixel intensity based on Eq. (1): Here, depicts the standardised intensity of a pixel; indicates the intensity value of a pixel; and refer to the minimal and maximal values of pixel intensity, respectively. Gamma correction Gamma correction is frequently employed to modify the diverse features in colour and brightness. The gamma coefficient is presented to depict the non-linear relations among the value of a pixel and its actual luminance23. The greater value of gamma and the steeper curve of these relations are affecting the contrast improvement. Now, signifies the correction result, denotes the original image, and regulates the level of correction. The values of increase the entire brightness, whereas improves contrast by darkening low-intensity regions. To select a proper gamma value on a per-image basis is vital, specifically for radiological datasets, where illumination and exposure differ substantially. Denoising via Bilateral Filter (BF) The BF employs a non-linear image denoise. It is advanced from the norm of spatial Gaussian convolution, which is linear and deliberates a single weighted factor to implement the smoothing of the image24. Equation (3) displays the principle of Gaussian convolution. where indicates outcomes of convolution, signifies the factor of normalisation, represents the original value of the pixel, and G implies Gaussian spatial convolution. BF acquires the principle and includes a factor based on intensity weighting. The principle of BF is expressed in Eq. (4). Now, denotes Gaussian intensity, and signifies Gaussian spatial convolution. In BF, the Gaussian spatial convolution has been biased alongside its weight intensities acquired from the adjacent pixels within the kernel. This deliberation offers a non-linear nature that generates an imagery with decreased noise, though maintaining edges. Figure 3 represents the pre-processed images. Pre-processed images. Feature representation strategy For feature representation, a fusion of three DL models, such as EfficientNet-B3, CapsNet, and ConvNext V2, is used to capture informative features crucial for classification efficiently. The EfficientNet-B3 model is effectual for its optimized and scalable convolutional feature extraction with robust parameter. Also, spatial hierarchies and relationships between features are efficiently captured by CapsNet, while also conserving part-whole relationships that conventional CNNs may overlook. Furthermore, modern convolutional design with transformer-inspired improvements is integrated by the ConvNext V2 method to improve feature robustness. In the proposed MDLF-EISD framework, multiple models are integrated to leverage complementary feature learning. EfficientNet-B3 extracts efficient multi-scale radiological features, CapsNet preserves spatial and part–whole relationships, and ConvNeXt V2 captures global contextual information. The fused features are further refined using a Bidirectional Long Short-Term Memory layer to model dependencies within the combined representations, improving feature robustness and classification performance. EfficientNet-B3 The EfficientNet employs a sophisticated CNN scaling model. A simple compound coefficient is leveraged to accomplish effective outcomes. Instead of traditional approaches, which result in a depth, width, and resolution of the network, EfficientNet constantly scales every dimension25. The outcome is enhanced when every size has been scaled; however, the execution is substantially improved when every characteristic of the network is balanced with respect to resource accessibility. Consequently, it improves the image resolution and coefficient of width and depth. The coefficient controls the admittance of new resources to scale the model. A primary grid , and searching can recognise these constants that describe how the added resources were assigned to width, resolution, and depth. EfficientNet-B3 EfficientNet is a family of CNN frameworks. EfficientNet‐B3 is among the family of EfficientNet and is designed to balance an effective approach. One of the vital novelties of EfficientNet‐B3 was leveraging a compound scaling approach by instinctively scaling up the structure of the method, concerning width and depth, depending on the resolution of the input image. The EfficientNet‐B3 is 3rd method in the family of EfficientNet, and contains a similar depth as the EfficientNet‐B0. Therefore, it comprises huge filters for each layer and maximal spatial resolution of mapping features. The method consists of stacked convolutional layers covered by a layer of batch normalisation (BN) and an activation function of ReLU. Feature maps utilise depth-wise separable convolution, which lessens the number of parameters and executes a single convolution per channel. This method could acquire more intricate representations in inputting image data, although upholding a moderately smaller set of parameters. In EfficientNet‐B3, lower computation is employed to accomplish improved outcomes. In contrast with B1, the B3 employs depth-wise separable convolutions that make it more effective and lead to a smaller model size and quicker inference. This method comprises a vast amount of data collected to train a model that is modified or employed. In this framework, 26 blocks of convolution were succeeded by a layer of convolution with activation of BN. An MB‐Convolution has a reversed residual block that contains a convolution layer, depth‐wise convolution, and a convolution layer alongside skipping connections at the start and end. Subsequently, the latent space dimensions were diminished through a global average pooling. CapsNet Despite the structure of CNN progress in numerous classification concerns, they hold multiple theoretical shortcomings: (1) the layers of max-pooling and average-pooling in CNNs contains few entities that the system finds to identify, (2) they necessitate wide data for learning effectually, and (3) it doesn’t taken into deliberation spatial hierarchic relationship among simpler and intricate images26. It upholds hierarchical spatial relations; this framework can learn quicker and employ fewer instances per class. In recent times, the effective utilisation of CapsNets incorporated an augmented model that uses iterative routing-by-agreement. It creates CapsNet, distinct from CNN, and this routing-by-agreement that offers it the capability of recognising the spatial relations. In CapsNet, layers of convolution eliminate lower-level aspects from images and send them to CapsNet layers comprising a vector of equal dimension. Here, the primary layer of capsule shapes must fit the outcome from the preceding layer of convolution. In Eq. (9), j denotes the capsule, and j depicts the squashing operation for the vector of input , and the vector of output . In a specific entity, the features in vector create diverse characteristics. Alike to classical CNN, the input has been weighted addition through every multiplication among coupling coefficient calculated by iterative “Routing by Agreement” method and prediction vector , here every prediction vector was formed through multiplication of outcome from a particular capsule in the preceding layer by trainable weighted matrix . From capsule i and capsule in the preceding layer j, the coupling coefficient was calculated by Softmax of to point out the likelihood that capsule i is united to capsule j. ConvNext V2 The ConvNeXt-V2 is a sophisticated framework of CNN integrating various novel ideas from ViTs. It tackles the variance among CNNs and ViTs. The objective of ConvNeXt-V2 is to uphold the efficacy and benefits of CNNs. Additionally, it offers numerous significant novelties that enhance the performance of the model27. The considerable progression is an integration of global response normalisation (GRN). GRN upgrades the model of global spatial dependency, though preserving local sensitivity. Consequently, it tackles a restriction viewed in standardisation models like LayerNorm and BatchNorm. It alters feature responses based on the global context. This increased the proficiency of method generalising in visually intricate settings. Specifically, the modification is vital to eliminate features and categorise silicosis that usually exhibit a wide variety of geometries, materials, and textures. Another substantial improvement in training models like stochastic depth, layer scaling, and large kernel depthwise convolutions. This improvement allows ConvNeXt V2 to accomplish a greater outcome, which is more effective than various transformer-based techniques. It exceeds its predecessor and matches or outperforms modern ViT approaches like Swin Transformers. The ConvNeXt V2 was employed for accomplishing higher execution, interpretability over attention features, and correspondence with convolutional inductive preferences. Specifically, these features are beneficial while working with structural data that entails constrained spatial hierarchy amid numerous kinds of silicosis. Feature fusion framework To attain a richer and extensive visual representation from radiological images, a hybrid multi-model feature fusion approach is utilised. The proposed method combines three heterogeneous feature extractors: EfficientNet-B3, CapsNet, and ConvNeXt V2, to acquire complementary hierarchical, spatial, and global contextual data from the input. Assume the pre-processed input image depicted as I. Every method independently eliminates a higher-dimensional feature embedding: Now, signifies multi-scale convolutional features formed by EfficientNet-B3, represents pose-aware, transformation-equivariant features acquired by CapsNet, acquires global response normalised attributes and long-range contextual dependency through ConvNeXt V2. To develop a unified representation, the individual embeddings were integrated utilising channel-wise concatenation: Since concatenation improves the entire dimensionality, a linear projection layer is employed for dimension and feature refinement. The fused vector feature is attained: Here, the W and implies learnable parameters of a fully connected projection layer. The succeeding fused representation combines local texture attributes, transform-equivariant capsule features, and global semantic context, offering strong, robust, and discriminative feature embedding for the downstream of radiological abnormality. Figure 4 shows the images of extracted feature maps. The Bright yellow/orange regions in the feature maps designate high activation where the model concentrates on significant structures like the lung fields, rib boundaries, and thoracic cavity. Dark purple areas signify low activation, representing areas with less relevance to the classification process. Extracted feature maps. Silicosis classification model A convolutional bidirectional attention mechanism is utilised for effective classification of silicosis into corresponding categories28. The local and global contextual data are effectively captured by the model while the bidirectional attention mechanism (AM) considers dependencies in various directions. The method also facilitates to concentrate on the most relevant areas, thus enhancing distinction between subtle disease phases. It also mitigates data loss and efficiently handles intrinsic patters. AM gives robustness, resulting in higher classification accuracy and more reliable predictions. CNN acquires local aspects of inputting data through a layer of convolutional, whereas the layer of pooling lessens the spatial dimension of mapping features, while retaining vital elements. The AM permits the method to arrange substantial data by allocating weights to every input channel. Because of its bi-directionality, the layer of Bi-LSTM could employ both forward and backward data for acquiring the longer‐term dependency of CNN mapping features. The CNN‐BiLSTM‐Attention model is capable of the task of image classification, utilising the proficiency of CNN, BiLSTM, and AM. This amalgamation allows it to effectually leverage regional and global aspects in the pre-processed image, while highlighting vital data. Consequently, it establishes sturdy performance through numerous challenges of image classification. CNN CNN contains the structure of a deep neural network, comprising convolutional, pooling, and fully connected (FC) layers. Convolutional operations were essential to the effectiveness of CNN. In a convolutional layer, the network evaluates input data throughout a convolutional operation, utilising a set of learnable filters called convolution kernels. Here, signifies the outcome mapping feature of layer i, represents the input of the convolutional layer, signifies the operation of dot product, implies the weighted matrix of convolution, f refers to the activation function, and indicates vector modification. Pooling operations were employed to lessen the spatial dimension of the mapping feature while maintaining vital aspects. Distinctive pooling operations comprise average pooling and max pooling, among others. These filters acquire numerous features when inputting data, like textures and edges. Now, indicates deviation, signifies the outcome of the maximal pooling layer, and represents the maximal pooling function. The mapping feature is attained throughout the pooling operation and is transmitted to the layer of FC that calculates the outcome vector. The FC layer endures at the apex of the network, transforms the features eliminated by previous pooling and the layer of convolutional into the outcome. Within this layer, each neuron was linked to every neuron in the last layer. Now, signifies the bias, denotes the weighted matrix, and represents the outcome vector. BiLSTM BiLSTM has been selected because of its capability to learn bidirectional contextual relationships between the fused features separated from EfficientNet-B3, Capsule Network, and ConvNextV2. In the proposed architecture, the fused feature representation is processed as a sequence of feature elements, allowing the BiLSTM model to acquire dependencies from both forward and backward directions. This ability improves global feature representation and improves classification performance. Additionally, BiLSTM provides an effective and computationally effective technique for modeling long-range feature relationships within the presented medical image classification architecture. BiLSTM includes a layer of forward LSTM and backward LSTM. Its effective sequence modelling capability depicts LSTM. To integrate memory units and gating modules, it can adeptly retain and convey longer-term dependency data, while discriminatingly forgetting useless data, consequently resolving the exploding and gradient vanishing issues challenged by conventional RNNs, allowing them to surpass intricate tasks, namely NLP. The gates and functions in LSTM neurons were evaluated: Now, Eq. (23) signifies the input gate employed to regulate and upgrade novel input data. indicates the outcome of the input gate; implies inputting data of the existing time step; refers to the hidden layer (HL) of the preceding time step; is the activation function of the sigmoid; and imply the weight and bias of the input gate, correspondingly. Equation (24) signifies the forget gate that regulates whether the memory of the preceding time-step has been forgotten; Eq. (25) refers to the cell-state upgraded by forget and input gates, and represents the candidate cell-state of the existing time-step. Equation (26) denotes the cell state upgraded by inputting, forget gates, and candidate cell state, where indicates the cell state of the existing time step. Equation (27) implies the output gate managing the outcome of HL of the existing time step. Equation (28) indicates assessing the HL of the existing time-step utilising the outputting gate and the upgraded cell state, where indicates the HL of the existing time-step. Now, t represents the counts of input order. To leverage the input order of 3 successive time-steps, thus taking t=3. X indicates a series of input orders, which means mapping the feature and AM. S implies a series of outcome orders succeeded by the succeeding layer of FC. U and D imply the backward and forward LSTM components, respectively. The structure of BiLSTM improves its proficiency to manage non-linear time-series, enhancing long-term temporal dependency and improving NN execution to accomplish precise outcomes. AM AM acquires vital data from a vast amount of data, enhancing the processing efficacy and outcome of perceptual details that upgrade the insight and deployment of essential data by allocating diverse weights to various aspects of input data; thus, it comprises more critical elements that have more impacts on outcomes. Meanwhile, the SE-AM have layers for scaling and dot multiplication, which can lessen the computational complexity of the method. The SE-AM comprises dual primary stages: Squeeze and Excitation. Here, H, W, and C imply height, width, and number of channels of the dataset, respectively; x denotes the original dataset, implies similar data. denotes some or a series of transformations. In the Squeeze step , the model condenses the input mapping feature U into a single value by a global average pooling operation to attain a vector of length C. This procedure is seen as compressing all data in the spatial dimension to a single point. In the Excitation stage , a layer FC is employed for mapping this vector to a lower dimension, which means a smaller vector, and the weights of every channel were standardised to standardise every component of this vector between zero and one using the sigmoid function. Eventually, these weights are multiplied by the input mapping feature to attain the weighted mapping feature and passed to the resultant layer of Bi-LSTM. Figure 5 denotes the block diagram of the convolutional bidirectional attention model. Block diagram of the convolutional bidirectional attention model. Assessment metrics Several performance measures are utilised to validate the model’s performance. In this study, accuracy, precision, recall, F1-Score, and AUC are employed. Accuracy assesses how close a measurement or prediction is to the true, which indicates its exactness or correctness. Precision is the percentage of true positive results among all positive classifications. It measures how many of the instances identified as positive were really correct. Recall is a measure that calculates the ratio of actual positive samples that are accurately predicted by a model. The F1-Score is the harmonic mean of a system’s precision as well as recall. It is utilised for balancing the importance of both FP and FN, which gives equal weight to precision and recall. The AUC score determines the system’s capability to differentiate between negative and positive classes across every possible classification threshold. Results interpretation The experimental analysis of the MDLF-EISD methodology is examined under Silicodata. The model is simulated using Python 3.6.5 on a PC with an i5-8600k, 250GB SSD, GeForce 1050Ti 4GB, 16GB RAM, and 1 TB HDD. The parameter setting is given as follows: learning rate: 0.01, dropout: 0.5, batch size: 5, epoch count: 300, and activation: ReLU. For experimental validation, 80:20 and 70:30 of training/testing dataset is used. Table 3 display the parameter setting of the proposed model. Moreover, these hyperparameters and model choices were selected to achieve a balance between accuracy, efficiency, and generalization. EfficientNet-B3 and ConvNeXt V2 are used as complementary feature extractors because they capture strong visual representations while remaining computationally efficient. Capsule Networks are included to better understand spatial relationships in the data, while attention-based fusion helps combine features from different models in a meaningful way. The convolutional bidirectional attention module further improves performance by capturing both local patterns and long-range dependencies. A moderately sized classification head with ReLU activation and Softmax ensures stable and accurate final predictions. Training is controlled using AdamW with a small learning rate and weight decay to improve generalization, while dropout and early stopping help prevent overfitting and ensure the model does not train beyond its optimal point. Table 3Parameter setting of proposed MDLF-EISD model.ModuleHyperparameterOptimal valueEfficientNet-B3 Feature ExtractorLearning Rate0.001Feature Dimension1536Dropout Rate0.2Capsule NetworkNumber of Capsules16Capsule Dimension16Routing Iterations3Hidden Units128ConvNeXt V2 Feature ExtractorNumber of Stages4Hidden Dimension192Drop Path Rate0.2Feature Dimension1024Feature Fusion ModuleFusion MethodAttention FusionFusion Dimension512Fusion Dropout0.2Convolutional Bidirectional Attention ModelNumber of Bi-Directional Layers2Hidden Units128Attention Dimension128Convolution Kernel Size3Dropout Rate0.2Classification LayerDense Layer Units256Activation FunctionReLUOutput LayerSoftmaxTraining ConfigurationOptimizerAdamWLearning Rate0.001Weight Decay1e-4Epochs50Early Stopping Patience10 Parameter setting of proposed MDLF-EISD model. Results analysis of proposed model Figure 6 presents confusion matrices of the MDLF-EISD model under 80% training and 20% testing phases. The result indicates that the MDLF-EISD approach has effective classification and detection of each class exactly. Confusion matrices of (a) 80% training and (b) 20% testing. Table 4; Fig. 7 depict the classifier outcome of the MDLF-EISD model under 80% Training Phase (TRPHE) and 20% Testing Phase (TSPHE). On 80% TRPHE, the Silicosis class attains , , , , and MCC of 98.60%, 97.30%, 98.69%, 97.99%, 98.62%, and 96.93%, respectively. Furthermore, the Tuberculosis class obtains , , , , and MCC of 99.14%, 98.48%, 97.99%, 98.23%, 98.75%, and 97.66%, respectively. On 20% TSPHE, the STB class reaches , , , and MCC of 98.03%, 94.87%, 97.37%, 96.10%, 97.81%, and 94.80%, respectively. Also, the Normal class accomplishes , , , , and MCC of 99.01%, 97.62%, 95.35%, 96.47%, 97.48%, and 95.91%, respectively. Table 4Classifier outcome of MDLF-EISD model under 80%:20%.ClassesAccuracyPrecisionRecallF-ScoreAUC ScoreMCCTraining Phase (80%)Silicosis98.6097.3098.6997.9998.6296.93STB98.1197.1195.1796.1397.1294.89Tuberculosis99.1498.4897.9998.2398.7597.66Normal99.0596.7397.4797.1098.4296.54Average98.7397.4197.3397.3698.2396.50Testing Phase (20%)Silicosis98.6897.2799.0798.1798.7797.15STB98.0394.8797.3796.1097.8194.80Tuberculosis99.01100.0096.1098.0198.0597.39Normal99.0197.6295.3596.4797.4895.91Average98.6897.4496.9797.1998.0396.31 Classifier outcome of MDLF-EISD model under 80%:20%. Classifier outcome of MDLF-EISD model under 80%:20%. Figure 8 depicts the average values of the MDLF-EISD model under 80% TRPHE and 20%TSPHE. On 80% TRPHE, the proposed MDLF-EISD model obtains average , , , , and MCC of 98.73%, 97.41%, 97.33%, 97.36%, 98.23%, and 96.50%, respectively. Besides, on 20% TSPHE, the proposed MDLF-EISD model gains average , , , , and MCC of 98.68%, 97.44%, 96.97%, 97.19%, 98.03%, and 96.31%, respectively. Average values of the MDLF-EISD model under 80%:20%. Figure 9 depicts the training (TRAN) and validation (VALD) accuracy of the MDLF-EISD technique on 80:20 over 300 epochs. Both curves constantly grow and slowly converge, which suggests that the method is learning efficiently. The VALD accuracy steadily remains slightly higher than the TRAN accuracy, implying that the technique is not overfitting and generalises well to unknown information. The fluctuation accuracy is expected because of the intricacy of the task, but the total increasing tendency reveals stronger performance and steadiness of the model. Accuracy curve of the MDLF-EISD algorithm on 80:20. Figure 10 shows the TRAN and VALD loss of the MDLF-EISD model on 80:20 over 300 epochs. Both curves exhibit a constant reducing tendency, suggesting that the method effectually minimises error in learning. The loss of VALD stays slightly lower than the training loss through most epochs, implying better generalisation and no signs of overfitting. Though some fluctuations are monitored, it has become steadier and more consistent. Loss curve of the MDLF-EISD algorithm on 80:20. Figure 11 displays confusion matrices of the MDLF-EISD model under 70% training and 30% testing phases. The result implies that the MDLF-EISD system has effective identification and recognition of every class accurately. Confusion matrices of (a) 70% training and (b) 30% testing. Table 5; Fig. 12 depict the classifier outcome of the MDLF-EISD model under 70% TRPHE and 30%TSPHE. On 70% TRPHE, the Silicosis class obtains , , , , and MCC of 97.84%, 97.44%, 96.40%, 96.91%, 97.51%, and 95.25% respectively. Furthermore, the Tuberculosis class gains , , , and MCC of 98.31%, 96.10%, 97.18%, 96.64%, 97.93%, and 95.51%, respectively. On 30% TSPHE, the STB class accomplishes , , , and MCC of 99.12%, 98.73%, 97.91%, 98.32%, 98.73%, and 97.73%, respectively. Also, the Normal class attains , , , and MCC of 98.46%, 96.50%, 93.88%, 95.17%, 96.61%, and 94.27%, respectively. Table 5Classifier outcome of MDLF-EISD model under 70%:30%.ClassesAccuracyPrecisionRecallF-ScoreAUC scoreMCCTraining Phase (70%)Silicosis97.8497.4496.4096.9197.5195.25STB99.1197.5098.8398.1699.0197.58Tuberculosis98.3196.1097.1896.6497.9395.51Normal98.7396.6695.2195.9397.3095.18Average98.5096.9296.9096.9197.9495.88Testing Phase (30%)Silicosis98.0397.0897.0897.0897.7995.59STB99.1298.7397.9198.3298.7397.73Tuberculosis98.4695.5498.1796.8398.3695.84Normal98.4696.5093.8895.1796.6194.27Average98.5296.9696.7696.8597.8795.86 Classifier outcome of MDLF-EISD model under 70%:30%. Classifier outcome of the MDLF-EISD model under 70%:30%. Figure 13 represents the average values of the MDLF-EISD model under 70% TRPHE and 30%TSPHE. On 70% TRPHE, the proposed MDLF-EISD model reaches average , , , , and MCC of 98.50%, 96.92%, 96.90%, 96.91%, 97.94%, and 95.88%, respectively. In addition, on 30% TSPHE, the proposed MDLF-EISD model accomplishes average , , , , and MCC of 98.52%, 96.96%, 96.76%, 96.85%, 97.87%, and 95.86%, respectively. Average values of the MDLF-EISD model under 70%:30%. Discussion Table 6; Fig. 14 demonstrate the comparative analysis of the MDLF-EISD model with other existing techniques under various metrics29,30. Depending on , the presented MDLF-EISD method has achieved superior of 98.73% whereas the SqueezNet, ViT, AlexNet, VGG16, ResNet50, Xception, and DenseNet121 models have obtained lesser of 70.00%, 60.00%, 92.62%, 93.03%, 92.76%, 97.29%, and 95.17%, correspondingly. Likewise, based on , the presented MDLF-EISD technique has a greater of 97.41% whereas the SqueezNet, ViT, AlexNet, VGG16, ResNet50, Xception, and DenseNet121 methodologies have attained a lower of 69.14%, 72.60%, 96.73%, 97.09%, 96.36%, 94.72%, and 95.48%, correspondingly. In addition, based on , the presented MDLF-EISD approach has accomplished a maximal of 97.36% whereas the SqueezNet, ViT, AlexNet, VGG16, ResNet50, Xception, and DenseNet121 techniques have gained a lower of 78.61%, 77.30%, 93.97%, 81.38%, 87.83%, 85.63%, and 96.01%, correspondingly. Furthermore, in accordance to the Computational Time (CT), the proposed MDLF-EISD model achieves a lesser CT of 3.22 s, compared to the other methods, such as SqueezNet, ViT Model, AlexNet, VGG16, ResNet50, Xception and DenseNet121, CT of 8.23 s, 5.10 s, 6.16 s, 8.44 s, 8.51 s, 5.99 s, and 4.56 s, respectively. Table 6Comparative analysis of the MDLF-EISD model with existing.Models MCCComputational time (sec)SqueezNet70.0069.1479.7778.6177.988.23ViT Model60.0072.6079.6277.3089.935.10AlexNet92.6296.7386.2093.9777.556.16VGG1693.0397.0986.8281.3884.868.44ResNet5092.7696.3686.8887.8381.358.51Xception97.2994.7294.1185.6377.675.99DenseNet12195.1795.4889.5396.0187.974.56MDLF-EISD98.7397.4197.3397.3696.503.22 Comparative analysis of the MDLF-EISD model with existing. Comparative analysis of the MDLF-EISD model with existing approaches. Table 7 depicts the computation efficiency of MDLF-EISD model with other techniques31. Based on inference time, the IAISAX-RTAD method provides lesser value of 2.88ms, Flops of 3.67G and GPU of 0.123 M, where the UwU-Net, UwU-Net + LeakyRelu, Small U-Net, Original U-Net, and ResNet18-U-Net Methods attain greater time of 6.25ms, 5.14ms, 4.59ms, 6.80ms and 6.51ms, respectively. Table 7Computation efficiency of the MDLF-EISD model.ModelFLOPs (G)Parameters (M)Inference time (ms)UwU-Net6.240.2636.25UwU-Net + LeakyRelu6.220.2635.14Small U-Net11.020.4674.59Original U-Net83.7931.0426.80ResNet18-U-Net8.1614.3216.51MDLF-EISD3.670.1232.88 Computation efficiency of the MDLF-EISD model. An ablation study is a model utilised in ML to assess the contribution of every factor of a system or model. In this process, specific features, parameters, or modules are altered or removed to monitor how performance variations occur. By comparing outcomes with and without particular factors, academic circles may classify which parts are most significant. These supports optimising the model and comprehending its internal actions. Table 8 shows the ablation study of the MDLF-EISD technique. The table values imply that the proposed MDLF-EISD (CNN-BiLSTM-Attention with feature fusion models) model has obtained higher , , of 98.73%, 97.41%, 97.33%, and 97.36%, respectively. Meanwhile, the other methods like CNN+EfficientNet-B3 (without BiLSTM and attention and Capsule network and ConvNext V2), CNN+capsule network (without BiLSTM and attention and EfficientNet-B3 and ConvNext V2), CNN+ConvNext V2 (without BiLSTM and attention and Capsule network and EfficientNet-B3), BiLSTM+EfficientNet-B3 (without CNN and attention and Capsule network and ConvNext V2), BiLSTM+capsule network (without CNN and attention and EfficientNet-B3 and ConvNext V2), and BiLSTM+ConvNext V2 (without CNN and attention and Capsule network and EfficientNet-B3) have attained minimal of 94.70%, 95.31%, 95.99%, 96.79%, 97.43%, and 98.18%, correspondingly. This detailed comparative analysis shows that the proposed MDLF-EISD model is effective and superior for the early identification of silicosis diagnosis. Table 8Ablation study of MDLF-EISD model.Models CNN+EfficientNet-B3 (without BiLSTM, attention, Capsule network, and ConvNext V2)94.7093.2793.2893.57CNN+capsule network (without BiLSTM, attention, EfficientNet-B3, and ConvNext V2)95.3194.0094.0194.11CNN+ConvNext V2 (without BiLSTM, attention, Capsule network, and EfficientNet-B3)95.9994.5294.6794.64BiLSTM+EfficientNet-B3 (without CNN, attention, Capsule network, and ConvNext V2)96.7995.2595.2895.40BiLSTM+capsule network (without CNN and attention and EfficientNet-B3 and ConvNext V2)97.4396.0496.0895.92BiLSTM+ConvNext V2 (without CNN, attention, Capsule network, and EfficientNet-B3)98.1896.7396.7696.57MDLF-EISD (CNN-BiLSTM-Attention with feature fusion models)98.7397.4197.3397.36 Ablation study of MDLF-EISD model. Figure 15 illustrate the Grad-CAM heatmaps highlighting the regions of chest X-ray images, which the model utilizes for prediction. Warmer colors (red and yellow) specify higher model attention, while cooler colors (blue and green) denote lower attention. In major cases, the model concentrates on the lung regions, signifying that it learns related pathological features for classification. Moreover, some activations appear outside the lung area, like near the shoulders or upper chest that can occur because of the surrounding anatomical structures or the coarse localization nature of Grad-CAM. Overall, the visualization helps interpret how the model identifies important regions in the images. Grad-CAM visualisation. Table 9 represents the performance of the K Fold Cross Validation across multiple folds. The CV 1 achieved accuracy of 86.31%, precision of 92.49%, recall of 87.78%, F1 score of 93.06%, and MCC of 85.76. CV 2 recorded accuracy of 89.53%, precision of 87.11%, recall of 90.92%, F1 score of 89.81%, and MCC of 86.39. CV 3 showed accuracy of 85.22%, precision of 91.30%, recall of 85.81%, F1 score of 94.39%, and MCC of 88.31. CV 8 reported accuracy of 90.13%, precision of 94.07%, recall of 86.67%, F1 score of 93.11%, and MCC of 93.05%. CV 9 demonstrated the highest accuracy of 95.80%, precision of 95.75%, recall of 94.92%, F1 score of 95.48%, and MCC of 95.04%. CV 10 maintained strong performance with accuracy of 93.64%, precision of 88.51%, recall of 91.41%, F1 score of 90.18%, and MCC of 95.65%, reflecting consistent generalization ability across unseen partitions. Table 9K-fold cross validation of MDLF-EISD model.K Fold Cross Validation MCCCV 186.3192.4987.7893.0685.76CV 289.5387.1190.9289.8186.39CV 385.2291.3085.8194.3988.31CV 491.6692.5585.4088.1588.99CV 587.5490.8590.8587.8293.89CV 690.1491.3290.9485.4391.84CV 788.3192.0885.2091.6889.08CV 890.1394.0786.6793.1193.05CV 995.8095.7594.9295.4895.04CV 1093.6488.5191.4190.1895.65 K-fold cross validation of MDLF-EISD model. The experimental outcomes illustrate the efficiency of the presented MDLF-EISD architecture for silicosis disease classification by chest X-ray images. The incorporation of EfficientNet-B3, Capsule Network, and ConvNextV2 allows the separation of rich and complementary feature representations on multiple views, thus enhancing the overall discriminative potential of the methodology. The feature fusion techniques additionally consolidate these representations into a specified feature space, improving inter-class separability. The stable performance recognized over various evaluation measures and various train-test split ratios, which illustrates the strongness of the presented architecture. However, the comparative analysis with existing methods and the ablation study validate the contribution of every individual component, confirming, which integrated influence of multi-backbone feature extraction, feature fusion, and attention-guided classification importantly improves silicosis recognition performance. Conclusion In this manuscript, a novel MDLF-EISD model has been designed to attain early and precise detection and classification of silicosis, thus facilitating timely clinical intervention and enhancing patient outcomes. To achieve that, the proposed MDLF-EISD model primarily pre-processed the input images through numerous steps: image resizing, pixel intensity normalisation, gamma correction, and denoising, which prepared the raw image data for analysis. For feature representation, a multimodal feature fusion approach is integrated, combining three DL architectures, namely EfficientNet-B3, CapsNet, and ConvNext V2, which effectively extract informative features essential for the classification process. Lastly, the convolutional bidirectional attention technique is employed for accurate classification of silicosis into corresponding categories. A comprehensive simulation investigation is conducted to guarantee the improved performance of the MDLF-EISD methodology. The comparative analysis of results inferred that the proposed MDLF-EISD model performed well compared to existing techniques across multiple evaluation measures. Thus, the proposed MDLF-EISD model is a promising solution for early and accurate identification and classification of silicosis using radiological images. The proposed system can be integrated into occupational health screening centres, enabling precise and consistent triage of CXR images, reducing dependence on radiologist interpretation variability. Traditional radiographic assessment of silicosis is limited by inter-observer variability. This model reduces classification uncertainty by utilising multimodal imaging features, improving consistency and precision. Even though the proposed MDLF-EISD model achieves a strong performance for silicosis classification from chest X-ray images, it has some limitations, this study is only focused on the single imaging modality (CXR) and depending upon on publicly available datasets, which may not fully denote the diversity of real-world clinical conditions. In addition, the methodology has not been extensively evaluated over multi-center datasets or various imaging devices, which might be influence its generalizability in broader clinical conditions. Future work should focus on integrating multimodal imaging, expanding geographic and demographic diversity, and conducting prospective clinical trials to validate model performance in real-world hospital environments. By integrating heterogeneous feature representations, the system reduces diagnostic ambiguity and provides more reliable precision diagnostics for high-risk occupational groups. This aligns with the goals of precision medicine by enabling targeted follow-up, early intervention, and individualised risk stratification. Deep features act like radiomic descriptors, and multimodal fusion captures both textural and spatial patterns, thus enhancing personalised precision in imaging-based classification. Acknowledgements This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea Government (MSIT) (No. RS-2024-00337489, Development of data drift management technology to overcome performance degradation of AI analysis models). Author contributions Conceptualization: Ramesh NSVSC SripadaData curation: Pothuri Surendra Varma, E. Laxmi Lydia, K. RadhikaFormal analysis: Ramesh NSVSC Sripada, E. Laxmi Lydia, K. RadhikaFunding acquisition: Cheolhee Yoon, Gyanendra Prasad JoshiInvestigation: Jang Mook Kang, Cheolhee Yoon, Gyanendra Prasad JoshiMethodology: Ramesh NSVSC Sripada, Pothuri Surendra Varma, E. Laxmi Lydia, K. RadhikaProject administration: Jang Mook Kang, Cheolhee Yoon, Gyanendra Prasad JoshiResources: Jang Mook Kang, Cheolhee YoonSoftware: Pothuri Surendra Varma, K. RadhikaSupervision: Cheolhee Yoon, Gyanendra Prasad JoshiValidation: Jang Mook Kang, Cheolhee Yoon, Gyanendra Prasad JoshiVisualization: Ramesh NSVSC Sripada, Pothuri Surendra VarmaWriting – original draft: Ramesh NSVSC SripadaWriting – review & editing: Gyanendra Prasad Joshi. Data availability The data that support the findings of this study are openly available at [https://iab-rubric.org/resources/healthcare-datasets/silico-data](https:/iab-rubric.org/resources/healthcare-datasets/silico-data) , reference number [22]. Declarations Competing interests The authors declare no competing interests. Ethical approval This article does not involve any studies with human participants or animals performed by any of the authors. Consent to participate Not applicable. Informed consent Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Gyanendra Prasad Joshi, Email: joshi@kangwon.ac.kr. Cheolhee Yoon, Email: bertter@police.ac.kr. References Associated Data Data Availability Statement The data that support the findings of this study are openly available at [https://iab-rubric.org/resources/healthcare-datasets/silico-data](https:/iab-rubric.org/resources/healthcare-datasets/silico-data) , reference number [22].