Comparing dermatologists' and artificial intelligence heat maps in dermoscopic image analysis via eye tracking

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• AI predictive systems use heat maps to highlight informative regions for diagnostic predictions.
• Comparing AI- and dermatologist-generated heat maps is crucial for assessing AI interpretability and clinical relevance.

Purpose of the Study:

• To compare dermatologists' and AI's heat maps generated from dermoscopic images.

Main Methods:

• Four dermatologists' eye movements were tracked while viewing 120 dermoscopic images (melanomas, basal cell carcinomas, squamous cell carcinomas, nevi, keratoses, vascular lesions).
• Class activation maps were generated using the DEXI algorithm for the same images.
• Overlap between dermatologist and AI heat maps was assessed using pixel-wise rank correlation.

Main Results:

• Median pixel-wise correlation between dermatologist and DEXI heat maps was 0.540.
• Inter-dermatologist correlation was 0.591, serving as an upper reference.
• Null correlations between DEXI and non-homologous dermatologist maps yielded a median of 0.434.

Conclusions:

• Substantial overlap exists between dermatologist and DEXI heat maps, suggesting shared diagnostic anchors.
• This overlap supports the potential interpretability of AI models in dermatology.
• Limitations include small sample sizes per lesion type and lack of lesion size data.