COLLECT: Counterfactual Oral Lesion Library for Explainable Concept Testing

Le résumé fourni par la source

This dataset contains clinical intra‑oral images collected by certified oral pathologists during routine dental examinations, using digital cameras under standard clinic lighting. Each image was verified by histopathology or expert assessment. Ethical approval was obtained from the McGill University IRB # [A07-M40-21B, 24-11-096]. The dataset includes four lesion types: (Aphthous Ulcer, Geographic Tongue, Hairy Tongue, and Oral Squamous Cell Carcinoma) selected from a larger study on interpretable AI for oral lesion classification. For each category, 10 high‑quality baseline images (40 total) were selected and validated by oral health experts. Each image was transformed using a structured counterfactual framework applying progressive size, color, and opacity modifications, producing five versions per image (original image and 4 counterfactuals) with a total of 600 images. Lesion size was either gradually increased or decreased to simulate growth or reduction; color was adjusted to mimic clinically similar lesion appearances; and opacity was reduced in steps to represent partial or complete lesion fading. These controlled variations simulate clinically meaningful changes and support research on AI model robustness, interpretability, and diagnostic reliability in oral imaging.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.