Reproducible Naevus Counts Using 3D Total Body Photography and Convolutional Neural Networks.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Naevus count is a key melanoma risk factor, but manual counting is inconsistent.
• Machine learning (ML) shows promise for dermatological image analysis.

Purpose of the Study:

• To evaluate automated, reproducible naevus counts using convolutional neural networks (CNNs) and 3D total body imaging.
• To assess the feasibility of ML in standardizing naevus counting for melanoma risk assessment.

Main Methods:

• Developed a CNN using 3D total body images from 82 subjects (57,742 lesions) for training.
• Validated the CNN on 10 subjects (4,868 lesions), with lesions classified by a senior dermatologist.
• Assessed CNN performance using sensitivity, specificity, and Cohen's kappa at lesion and person levels.

Main Results:

• Lesion-level analysis showed 79% sensitivity and 91% specificity for naevi ≥2 mm.
• Substantial agreement (Cohen's kappa 0.72) was found for naevi ≥5 mm.
• Person-level agreement reached 70% between automated and gold standard counts.

Conclusions:

• Automated naevus counting via 3D imaging and CNNs is feasible and achieves reasonable agreement with expert assessment.
• This approach offers a potentially faster and more reproducible alternative to manual naevus counting.
• The algorithm may improve consistency in melanoma risk stratification.