A Lightweight Attention-Guided Multi-Scale Ensemble Framework for Robust Skin Disease Classification

Resumo

Early treatment and minimization of the impact of the dangerous complications, such as skin cancer, is impossible without adequate diagnosis of skin disease. Nevertheless, the commonly used traditional diagnostic methods rely on the visual inspection of dermatologists, and thus the procedure is subjective, time-consuming, and unavailable in remote and resource-limited environments. To address these drawbacks, the following paper presents a small and multi-stage deep learning architecture that can be used to classify intelligent skin diseases with a balance between significant accuracy and high computing efficiency. The given system is expected to be efficient in terms of processing the dermoscopic images and is still applicable to being implemented on the devices with limited computer capabilities. The framework starts with the Adaptive Skin Image Noise Reduction Module (ASINRM) that improves image quality by using illumination normalization, noise reduction, contrast enhancement and intensity standardization. The step of preprocessing minimizes variability brought by the lighting conditions and imaging artifacts. The processed images will then be fed to the Multi-Scale Residual Feature Extraction Network (MSRFEN) to undergo accurate lesion segmentation to remove the affected area and isolate it with the rest of the healthy skin. This step of segmentation enhances localization of features and makes the later learning to be centered on clinically relevant areas. MSRFEN also simultaneously derives discriminative lesion detail at several spatial scales, including both fine-grained texture and coarse detail structural features, required to classify to high accuracy.  The next is the Cross-Scale Adaptive Feature Fusion Network (CSAFFN) which combines the multi-resolution features via an attention-guided fusion mechanism which focuses on the most informative representations. To classify, the Lightweight Ensemble Skin Disease Classification Module (LESDCM) employs several effective learners which work in parallel to produce solid probability predictions. Lastly, the Skin Lesion Robust Ensemble Prediction Aggregation (REPASL) module is an ensemble-based prediction aggregation module that is adaptive and uses confidence-based weighting and consensus boosting techniques to deliver a dependable final diagnosis. It is experimentally proven that the suggested framework has high classification accuracy, high generalization performance, and low computational complexity, which makes it appropriate to be applied in the real-life clinical and mobile healthcare setting.