Advancing skin cancer diagnosis with a multi‐branch ShuffleNet architecture

Abstract

AbstractIn this study, we present an innovative approach for enhancing skin cancer classification through a multi‐branch architecture inspired by ShuffleNet. Our methodology focuses on improving feature extraction and representation, emphasizing cross‐channel information exchange to achieve superior accuracy. The architecture comprises three branches: a primary feature enhancement branch, a parallel feature enhancement branch, and a global feature aggregation branch. We employ transposed convolution layers for upsampling, cross‐channel normalization, and the Swish activation function to enhance feature representations. Channel shuffle operations and group convolutions stimulate cross‐channel information exchange in both branches. The global feature aggregation branch utilizes global average pooling and depth concatenation to combine features from all branches. Subsequent Swish activation, followed by fully connected layers and softmax activation, yields class probabilities for precise skin cancer classification. This multi‐branch framework offers a promising avenue for accurate and informed medical image analysis in skin cancer detection. Integral to this research is the utilization of the ISIC2019 and ISIC 2020 datasets, encompassing diverse dermoscopic images from multiple sources. By leveraging these datasets, the methodology capitalizes on comprehensive data for precise skin cancer detection, thereby advancing medical image analysis.