Advancing skin cancer diagnosis with deep learning and attention mechanisms

Skin cancer, particularly melanoma, remains one of the most lethal diseases globally due to challenges in early detection and diagnosis. Conventional image segmentation models often face difficulties due to the high variability in lesion appearance and their limited ability to focus on critical features, thereby compromising diagnostic accuracy. In this study, we introduce an advanced AI-driven framework that integrates a Scaled Dot Attention Mechanism (SDAM) with a modified UNet architecture to improve skin lesion detection. The SDAM, applied as an attention mechanism between the encoder and decoder stages of the UNet, allows the model to prioritize relevant lesion areas and extract essential features while reducing noise. We evaluate the proposed model using the HAM10000 dataset, a diverse collection of skin lesion images, and test it on two additional datasets: ISIC (Preliminary) and PH2 (Preliminary), to assess generalization across various skin lesion types. Our model achieves significant improvements in melanoma detection with Dice scores between 0.97 and 0.988, accuracy ranging from 97.8% to 98.3%, and substantial enhancements in sensitivity. These results outperform baseline models, including standard UNet (Dice score: 0.85, accuracy: 88.4%) and DenseNet (Dice score: 0.87, accuracy: 90.1%). Furthermore, the model's performance was compared to state-of-the-art methods such as Attention UNet, UNet++, and TransUNet, consistently demonstrating superior results. Statistical analysis via a paired t-test reveals a significant performance boost (p-value = 0.02), further validating the effectiveness of the SDAM-enhanced approach. These findings highlight the potential of AI in advancing early skin cancer detection and diagnosis, with the SDAM-UNet framework offering prospects for personalized care and real-time clinical integration. Additionally, our model's performance across multiple metrics such as precision, recall, F1-score, and IoU showcases its robustness in classifying both melanoma and benign skin lesions, reinforcing its utility in clinical practice.