An online cutaneous melanoma risk screen tool

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Cutaneous melanoma detection relies on accurate diagnosis, with missed cases leading to poorer outcomes.
• Bridging the gap between AI research and clinical application is crucial for effective tools.
• Existing tools often use dermoscopic images, limiting applicability in primary care settings.

Purpose of the Study:

• To develop and deploy a publicly accessible online risk estimation tool for cutaneous melanoma.
• Prioritize high recall to minimize missed melanoma diagnoses using clinical photographs.
• Implement an AI model trained on real-world clinical images for practical screening.

Main Methods:

• A 'segmentation-first, then-classification' strategy was employed using MFSNet and a customized EfficientFormerV2-L model.
• A balanced dataset of 934 clinical images (melanoma and nevus) was utilized.
• Weighted loss function and external validation on smartphone images were incorporated.

Main Results:

• The customized model achieved a recall of 0.95, precision of 0.75, and F1 score of 0.84.
• Dermoscopic pretraining unexpectedly reduced classification performance on clinical images.
• External validation showed significantly higher malignant potential scores for melanoma vs. nevus (P < 0.0001).

Conclusions:

• A practical, high-recall online melanoma screening tool using clinical images was developed and deployed.
• The tool's free public access addresses critical gaps in early melanoma detection.
• The emphasis on minimizing false negatives supports improved outcomes, especially in primary care.