🔄 First Visit: Redirecting automatically to original journal site...
⏸️ STOP REDIRECT & STAY ON PAGETOWARD EXPLAINABLE AND RESOURCE-EFFICIENT MEDICAL IMAGE CLASSIFICATION A CRITICAL REVIEW AND PROPOSED HYBRID CNN–TRANSFORMER WEB FRAMEWORK
Abstract
Abstract
Despite the significant progress made by deep learning in image classification, there are four common problems that limit the potential of practical deployment of medical image classifiers: convolutional neural networks (CNNs) excel at modelling locally available texture but provide weak global context; vision transformers offer strong global context but require a lot of data and computation; many state-of-the-art classifiers are barely interpretable; and research prototypes have not been transferred to secure, responsive web applications. This paper reviews and summarizes the relevant literature that was identified in a BSCS image-classification project, and it adds basic literature on residual networks, lightweight CNNs, hierarchical transformers, multiscale attention, and gradient-based explanations, which are all technically relevant. Based on this synthesis, it introduces a resource-efficient hybrid system designed to construct multi-resolution feature maps with a lightweight pre-trained CNN, to model cross-region dependencies in a compact multiscale transformer, to combine the features for classification in a gated manner, and to synthesize complementary explanations using Grad-CAM and attention-rollout. There is a web deployment architecture defined, with image validation, privacy-sensitive pre-processing, model inference, calibrated confidence reporting, explanation overlays, and auditing. A stratified cross-validation protocol, imbalance-aware metrics, calibration, latency, model size, ablation studies and statistical comparisons with CNN and transformer baselines are also included in the paper, as they are reproducible. It is not just a precision claim that will come to the fore, but a coherent and actionable vision that connects model design, explainability, resource efficiency, and web deployment. The framework aims to facilitate future empirical application to a well-documented medical image data set, and the achievement of a closer link between classification research and deployable software for aiding in decision making.
Despite the significant progress made by deep learning in image classification, there are four common problems that limit the potential of practical deployment of medical image classifiers: convolutional neural networks (CNNs) excel at modelling locally available texture but provide weak global context; vision transformers offer strong global context but require a lot of data and computation; many state-of-the-art classifiers are barely interpretable; and research prototypes have not been transferred to secure, responsive web applications. This paper reviews and summarizes the relevant literature that was identified in a BSCS image-classification project, and it adds basic literature on residual networks, lightweight CNNs, hierarchical transformers, multiscale attention, and gradient-based explanations, which are all technically relevant. Based on this synthesis, it introduces a resource-efficient hybrid system designed to construct multi-resolution feature maps with a lightweight pre-trained CNN, to model cross-region dependencies in a compact multiscale transformer, to combine the features for classification in a gated manner, and to synthesize complementary explanations using Grad-CAM and attention-rollout. There is a web deployment architecture defined, with image validation, privacy-sensitive pre-processing, model inference, calibrated confidence reporting, explanation overlays, and auditing. A stratified cross-validation protocol, imbalance-aware metrics, calibration, latency, model size, ablation studies and statistical comparisons with CNN and transformer baselines are also included in the paper, as they are reproducible. It is not just a precision claim that will come to the fore, but a coherent and actionable vision that connects model design, explainability, resource efficiency, and web deployment. The framework aims to facilitate future empirical application to a well-documented medical image data set, and the achievement of a closer link between classification research and deployable software for aiding in decision making.