Melanoma Detection with MobileNetV2 and Channel Attention: Addressing the Diverse Skin Tone Challenge

Skin cancer detection, particularly melanoma, is crucial for early treatment. However, datasets predominantly feature white skin tones, limiting generalization. To address this, datasets were adjusted to include brown skin tones. This study proposes a MobileNetV2-based model enhanced with dilated convolution and channel attention to improve classification across diverse skin tones. The model achieved superior performance on the modified ISIC-17 with an F1 score of 0.88, precision of 0.88, and recall of 0.87, and on modified UMCG with an accuracy of 0.85, precision of 0.84, and recall of 0.87. Ablation studies confirmed the significance of channel attention and dilated convolution in improving performance. These results emphasize the importance of diverse datasets and architectural innovations for robust melanoma detection.