AI-Powered Skin Cancer Detection with Dual-Phase Deep Learning, Clinical Fusion and Explainable AI

Skin cancer is among the most common cancers worldwide, and melanoma found late has high mortality, yet dermatologists are scarce in rural and low-resource regions. This project, “AI-Powered Skin Cancer Detection with Dual-Phase Deep Learning, Clinical Fusion and Explainable AI”, builds a decision-support system that works from smartphone photographs and a short clinical questionnaire instead of dermoscopy equipment. Phase 1 is an EfficientNet-B3 classifier trained on ISIC 2019 (8 lesion classes, lesion-level split). Phase 2 is a two-head fusion model that joins the image embedding with 14 clinical features and is fine-tuned on the smartphone dataset PAD-UFES-20 (6 classes, patient-level split). A rule-based red-flag layer covers six rare cancers that have no public training images. Grad-CAM, ABCD dermoscopy scoring and SHAP explain each prediction. Phase 1 reached 0.647 ± 0.022 balanced accuracy and 0.579 ± 0.020 macro F1 in 5-fold cross-validation. Phase 2 reached 0.689 macro F1 and 0.704 balanced accuracy with clinical fusion, against 0.694 and 0.736 for the image-only head, so fusion did not improve accuracy. A fairness audit shows a 0.247 sensitivity gap between light and medium skin tones on DDI, and no malignant case detected in the DermaCon-IN neoplasm subset (52 images, Fitzpatrick III-V only, no light-skin baseline). The system is intended to support triage and referral, not to diagnose.