AI‐assisted diagnosis of nail unit melanoma and melanonychia using a clinical deep learning model

BACKGROUND AND OBJECTIVES: Nail unit melanoma (NUM) is a rare but potentially fatal malignancy often misdiagnosed as melanonychia. Because biopsy may cause permanent nail dystrophy, accurate noninvasive diagnosis is essential. This study aimed to develop and validate an artificial intelligence model to distinguish NUM from benign melanonychia using clinical images and to assess its diagnostic utility through human comparison and external validation. PATIENTS AND METHODS: Clinical images from 172 patients with melanonychia and 122 patients with NUM were retrospectively collected. Pediatric melanonychia was excluded to minimize diagnostic ambiguity. Three convolutional neural networks (CNNs) were trained and validated with patient-wise splits. Human validation involved 40 representative images assessed by dermatologists, residents, and non-medical participants, with and without CNN assistance. RESULTS: ResNet-50 achieved the highest sensitivity (87.2 %). DenseNet-121 achieved the highest AUROC (area under the receiver operating characteristic curve) (0.954). CNN assistance improved the diagnostic accuracy of all human raters (accuracy increased from 70.0 % to 80.8 %), with dermatology residents showing the largest gain. CONCLUSIONS: The proposed CNN-based models demonstrated robust performance in differentiating NUM from melanonychia and improved diagnostic accuracy and agreement among human evaluators. While not intended to replace clinical judgment, this approach shows promise as a supportive screening tool in clinical settings, warranting further validation in larger, multi-institutional, and multi-ethnic cohorts.