A Hybrid Deep Learning Approach for Skin Cancer Classification Using Swin Transformer and Dense Group Shuffle Non-Local Attention Network

Abstract

Skin lesions are localized regions of the skin exhibiting atypical growth or appearance and can be either benign or malignant. Detecting and classifying these lesions accurately is crucial for early diagnosis and effective treatment of skin cancer. However, existing approaches for classifying cancerous and non-cancerous skin lesions from dermoscopic images are limited in capturing the appropriate features and also susceptible to overfitting in imbalanced datasets. This work aims to improve skin cancer classification by enhancing feature representation in dermoscopic images. A hybrid classification system is proposed, combining two tracks: the Swin Transformer and the Dense Group Shuffle Non-Local Attention (DGSNLA) Network, which integrates DenseNet169, Group Shuffle Depth-wise blocks (GSDW), and an enhanced non-local attention block (ENLA). This deep feature fusion approach effectively aggregates global and local features, resulting in more accurate feature representation and improved classification performance. To the best of our knowledge, this is the first study to combine a transformer with a custom CNN for skin cancer classification. The proposed network achieves an accuracy of 94.21%, an F1-score of 96.64%. a precision of 96.62%, and a recall of 96.25% on the HAM10000 dataset.

Keywords

• Convolutional neural network
• deep learning
• dermoscopic images
• image processing
• skin cancer