Attention-Based Inception-Residual CNN: Skin Cancer Diagnosis with Attention-Based Inception-Residual CNN Model.

Abstract

Skin cancer poses a significant global health concern, demanding early diagnosis to enhance patient outcomes and alleviate healthcare burdens. Despite advancements in automated diagnosis systems, most existing approaches primarily address binary classification, with limited focus on distinguishing among multiple skin cancer classes. Multiclass classification poses significant challenges due to intra-class variations and inter-class similarities, often leading to misclassification. These issues stem from subtle differences between skin cancer types and shared features across various classes. This paper proposes an attention-based Inception-Residual CNN (AIR-CNN) model specially designed to tackle the challenges related to multiclass skin cancer classification. Incorporating the attention mechanism model effectively focuses on the most relevant features, enhancing its ability to distinguish between visually similar classes and those with intra-class variations. The attention mechanism also facilitates effective training with limited samples. The inclusion of Inception-Residual (IR) blocks mitigates vanishing gradients, improves multi-scale feature extraction, and reduces parameters, creating a lightweight yet accurate model. The experimental evaluation of the ISIC 2019 dataset demonstrates superior performance with 91.63% accuracy and fewer parameters than state-of-the-art methods, which makes it suitable for practical applications, thus contributing to the advancement of automated skin cancer diagnosis systems.