SE-VGG16: Enhanced classification of skin cancer lesions using VGG16 backbone with squeeze-and-excitation blocks

Skin cancer is a global health concern that must be addressed. Non-melanoma skin cancer (NMSC), in the forms of basal cell carcinoma (BCC) and squamous cell carcinoma (SCC), is the most common malignancy in the world with instances outstripping all other cancers put together. Existing evidence shows a spectacular escalation in NMSC incidence, mostly among the elderly. Conversely, malignant melanoma, while less common but more lethal due to its high metastatic potential, is the cause of most skin cancer fatalities. SE-VGG16 is proposed in this study as a model that combines Squeeze-and-Excitation blocks with VGG16 structure for classifying lesions of skin cancer. The model is verified by testing with the ISIC 2019 dataset downsampled to a binary ‘Cancer’ and ‘Non-Cancer’ classifier. With the incorporation of channel-wise attention mechanism, SE-VGG16 achieved an overall accuracy of 81.0% which was higher compared to that of conventional VGG16 (80.6%), Xception (80.5%), ResNet50 (67.2%), and ConvNeXt-Base (68.3%). Significant attention was highlighted in non-cancer classification metrics with precision score of 80.4%, recall score of 84.7%, and F1-score of 82.5%. The model was found to be able to yield low false positive rates along with good interpretability and hence was a good contender for decision support systems in diagnosing skin cancer.