Skin Cancer Diagnostics AI

Abstract

Skin lesions can go unnoticed by non-experts, and diagnosis relies on visual assessment and biopsy, leading to delays in care. This paper develops and evaluates a deep learning approach for a three-class (malignant, benign, and non-neoplastic) skin lesion classification. The Fitzpatrick17k dataset was augmented to 21, 864 labelled images, with an approximate train-test-validation split of 85-10-5. The model uses an EfficientNetV2-S pretrained on ImageNet, swapping the final layer for three-class prediction. The model was trained using cross-entropy loss and AdamW, with separate learning rates for the backbone and classifier head. To improve robustness and reduce overfitting, data augmentation techniques, MixUp, and automatic mixed precision were applied. The best-performing model achieved 91.37% overall accuracy, macro F1-score of 87.53%, and malignant recall of 89.05%. These results suggest that the proposed method could support early triage and lesion screening for non-experts.