Skin Cancer Detection in 3-Dimensional Total Body Photography (3D-TBP) using Deep Learning

Early detection is critically important for the best outcomes in treating skin cancer, especially aggressive melanoma. An automated skin cancer classification system employing a Custom Convolutional Neural Network (CNN) in 3D Total Body Photography (3D-TBP) to classify lesions as benign or malignant is presented in this paper. Images of the same resolution are preprocessed with rigorous preprocessing including watershed segmentation for lesion isolation, resizing, and normalization to focus on cancerous areas. Advanced techniques of data augmentation guarantee improved model generalization on different scenarios. Hierarchical feature extraction blocks optimized at extracting intricate patterns of skin lesions from the heart of the CNN architecture. Efficient convergence and robust performance are obtained by training with the AdamW optimizer and using the OneCycleLR scheduler. The system achieves an overall accuracy of 74.64% with high specificity for malignant cases. Real time classification results and confidence scores along with the visual segmentation output is deployed via Streamlit, a web application, accessible to all users. This comprehensive strategy tackles important issues encountered in automated skin cancer detection and allows for a scalable, user-friendly diagnostic procedure and prepares the basis for additional refinement in terms of sensitivity, dataset diversity and explainable AI.