SkinSight: Automated Skin Cancer Detection Using Deep Learning

Abstract

Skin cancer is a rising global health problem in which early diagnosis can significantly improve the effectiveness of therapy. In this project, a dual-model deep learning approach for skin cancer diagnosis on Total Body Photography (TBP) and dermoscopic images is presented. Two separate convolutional neural networks based on EfficientNet-B0 were trained and compared using the ISIC 2024 and ISIC 2019 datasets. The TBP model was also trained for binary prediction (benign or malignant) from cropped TBP images of ISIC 2024. The dataset was additionally balanced to address class imbalance and Focal Loss was employed to increase the sensitivity of the model towards malignant cases. The final model showed an accuracy of 86.71%, precision of 84.52%, recall of 89.87%, F1-score of 87.12%, and an AUC of 0.9313, which proves its high clinical reliability. For dermoscopic images, a multi-class classifier was built based on EfficientNet-B0 augmented with a Channel Attention mechanism for enhanced fine-grained feature capture for eight types of lesions in the ISIC 2019 dataset. CLAHE contrast enhancement and removal of hair artifacts were used for preprocessing. The model resulted in 67.10% accuracy, F1-score of 67.01%, and average AUC of 0.9139 with superb per-class performance for vascular and dermatofibroma lesions. The system is implemented with a Flask backend that serves the trained models as REST APIs and a React.js frontend to engage and visualize real-time prediction. The results indicate that the dual-model architecture works well for different types of clinical images and is a viable AI assisted diagnostic tool for skin cancer screening.