Multidimensional CapsNets attention-gated approach for skin cancer detection and classification

Abstract

Skin cancer remains a major global cause of mortality, and early detection in its premalignant stages is crucial for improving patient outcomes. Traditional diagnostic methods face challenges such as time-consuming analysis, and limited accuracy. This study introduces the Multidimensional Capsule Networks Attention-Gated Module (MCAGM), an advanced automated deep learning framework designed to overcome these limitations. The MCAGM model utilizes Capsule Networks (CapsNets) enhanced with an pioneering spatial-channel attention mechanism, specifically designed to highlight clinically significant features in dermoscopic images (HAM10000 dataset) while effectively suppressing noise. The dual-domain attention mechanism (spatial and channel) dynamically refines feature importance, eliminating subjective interpretation and ensuring objective prioritization of relevant features. This end-to-end automated system dramatically reduces diagnosis time from hours to seconds, offering a significant improvement in efficiency. Furthermore, the CapsNet-based spatial hierarchies preserve critical lesion patterns that are often missed by conventional Convolutional Neural Networks (CNNs), enhancing the model's ability to detect subtle features and improve diagnostic accuracy. The model achieves exceptional performance with 97.63 % accuracy, 98.11 % precision, and 98.73 % recall, outperforming state-of-the-art methods by 8–19 % in accuracy (e.g., CNN: 88.88 %, CapsNet: 86.84 %), demonstrating its potential as a reliable tool for skin cancer diagnosis.