A Robust Deep Learning Framework for Multispectral Medical Image Fusion and Diagnosis for Skin Cancer

Abstract

Skin cancer has become a serious worldwide health issue which requires doctors to make quick and correct skin cancer diagnoses because this process determines how successful treatments will be and how long patients will live. The process of accurate identification of skin lesions remains difficult for dermatologists because benign and malignant skin lesions share common visual characteristics. To solve this problem we present SkinNetX which is an innovative deep learning system that combines ConvNeXtV2 block. The SkinNetX system uses its architectural design to achieve two objectives which are to gather detailed information and to boost its ability to differentiate between different things. The system uses ConvNeXtV2 blocks at its beginning stage to identify two different types of patterns which include detailed local patterns and larger regional texture patterns that help differentiate between two similar types of lesions. The system uses a split self-attention mechanism in its next stage to identify important areas of lesions which helps patients understand the process while the system achieves better performance than standard self-attention systems. The increasing worldwide incidence of skin cancer needs effective diagnosis methods which deliver precise results because they determine treatment success and patient survival rates. The classification process becomes difficult because expert dermatologists cannot differentiate between malignant and benign skin lesions which share identical visual characteristics. Our team developed SkinNetX as a deep learning framework which combines ConvNeXtV2 blocks with a dissociated self-attention method to achieve accurate and fast skin lesion development. The SkinNetX design uses its architecture for two purposes which include gathering detailed information and improving learning of distinct features. The initial phases use ConvNeXtV2 blocks to identify both minute local details and broader texture elements which help in differentiating between closely resembling lesion types. The subsequent stages use split self-attention which enables the system to concentrate on important clinical lesion sections while decreasing the need for processing power that comes with standard self-attention techniques.