Small lesion, big challenge? Diagnosis of small diameter melanocytic lesions by dermatologists using AI support

Background: Dermoscopy enhances melanoma detection, but small-diameter melanomas (SDMs) remain diagnostically challenging. Convolutional neural networks (CNNs) may detect subtle patterns beyond human perception. This study evaluates how lesion diameter influences the diagnostic accuracy of dermatologists, with and without CNN support. Patients and methods: This multicenter cross-sectional study included 150 histopathologically verified lesions: 110 small lesions (70 SDM, 40 nevi) and 40 large-diameter melanomas (LDM > 5 mm). Dermatologists evaluated lesions at level: (I) dermoscopy, close-up image, and clinical metadata; (II) plus CNN predictions. Primary outcomes were sensitivity, specificity, and ROC-AUC. Results: The CNN achieved a sensitivity of 62.7% (53.4%–71.2%), specificity of 80.0% (65.2%–89.5%), and ROC-AUC of 0.800 (0.723–0.876). Dermatologists performed higher at 71.8% (62.8%–79.4%), 82.5% (68.1%–91.3%), and 0.853 (0.794–0.912), improving slightly with CNN support to 72.7% (63.7%–80.2%), 85.0% (70.9%–92.9%), and 0.860 (0.796–0.923) (all p > 0.180). Subgroup analyses showed lower CNN performance in SDM, with sensitivity 48.6% (37.3%–60.1%) and ROC-AUC 0.740 (0.644–0.836), compared with LDM at 87.2% (73.3%–94.4%) and 0.904 (0.836–0.973). Dermatologists also reached their highest ROC-AUC in the LDM subgroup (0.925 [0.858–0.991]). Overall, CNN support led to slight, non-significant improvements, most notably in SDM (all p > 0.720). Conclusions: Lesion size strongly influences melanoma diagnosis. CNN support offers modest, non-significant gains, with highest accuracy achieved by experts using AI assistance.