Influence of hair presence on dermoscopic image analysis by AI in skin lesion diagnosis.

The escalating rates of melanoma and non-melanoma skin cancers highlight the urgent need for enhanced diagnostic tools. This study evaluates the effect of hair presence in dermoscopic images on AI-driven skin lesion recognition, using 10,015 images from the HAM10000 collection. Images were categorized based on the extent of hair occlusions to assess their impact on AI performance. Advanced metrics like Gradient-weighted Class Activation Mapping and novel infection metrics provided deeper insights into the model's focus and segmentation efficacy. The results revealed that the AI model achieved an impressive overall accuracy of 95.3 % and an Area Under the Curve (AUC) of 99.1 %, maintaining high performance across various lesion types. Performance varied with hair occlusion levels; it was optimal under few hair occlusions but dropped significantly when hair was abundant. For instance, Vascular Lesions saw a dramatic decrease in performance metrics from 0.515 to 0.115 under heavy hair occlusion, while Actinic Keratoses and Intraepithelial Carcinoma were least affected. Sparse hair occasionally improved accuracy, suggesting its potential utility in training robust models. The study underscores the complex impact of hair on AI diagnostics in dermatology, advocating for the development of advanced preprocessing techniques and hair-robust algorithms to enhance AI diagnostic capabilities.