The number of nevi on a person is the most potent risk factor for melanoma; however, its counting is highly inconsistent due to a lack of even methodology and inter-rater agreement. Machine learning emerges as a valuable tool for image classification in dermatology. 

A recent study tested the possibility of automated, reproducible nevus counts by merging convolutional neural networks (CNN) and three-dimensional (3D) total body imaging. 

The study used total body images from an examination of nevi in the general population for the training (82 subjects; 57,742 lesions) and testing (10 subjects; 4,868 lesions) datasets for developing a CNN. A senior dermatologist labeled the lesions as nevi or not ("non-nevi") as the gold standard. The performance of the CNN was assessed using sensitivity, specificity and Cohen′s kappa and evaluated at the lesion level and person level. 

• Lesion-level analysis that compared the automated counts to the gold standard showed a sensitivity and specificity of 79% and 91%, respectively, for lesions ≥2 mm, and 84% and 91%, respectively, for lesions ≥5 mm.
• Cohen′s kappa of 0.56 indicated a moderate agreement for nevi ≥2 mm, and substantial agreement (0.72) for nevi ≥5 mm.
• A person-level agreement found 70% agreement between automated and gold standard counts for a set of 10 test participants.
• However, subjects with numerous seborrheic keratoses showed lower agreement. 

Thus, the combination of 3D total body photography and CNN can enable automated nevus counts with a reasonable agreement, similar to an expert clinician. 

Such an algorithm can be a faster, reproducible method than traditional in-person total body nevus counts.

Source: Betz-Stablein B, D’Alessandro B, Koh U, et al. Reproducible naevus counts using 3D total body photography and convolutional neural networks. Dermatology. 2022;238(1):4-11.