Factors to consider in a primary care setting when using convolutional neural network tools for melanoma diagnostics: a retrospective analysis of images and patient characteristics.

Primary care physicians typically perform the initial assessment of skin lesions. Artificial intelligence tools, using techniques such as convolutional neural networks, are under development and have shown promise for evaluating dermoscopic images for the detection of skin cancer. These tools could possibly support physicians. This study investigated possible factors that might affect the specificity of an artificial intelligence tool, associated with the patient, lesion, or dermoscopic characteristics, which clinicians could potentially avoid, or for which the tool could be optimized. A total of 228 dermoscopic images from a prospective clinical trial in primary care were retrospectively examined by a resident general practitioner using conventional dermoscopy algorithms to describe the lesion characteristics. All lesions falsely classified as melanomas suspected by the artificial intelligence tool in the trial were compared with the true negative lesions with regard to lesion and patient characteristics. Elevated lesions were more often correctly classified as benign compared with macular lesions [odds ratio (OR) = 16, 95% confidence interval (CI) = 2.0-120, P  = 0.008]. The tool was more likely to falsely suspect melanoma if the ruler in the dermoscopic image was located over the lesion (OR = 2.0, 95% CI = 1.1-3.4, P  = 0.017). This indicates that elevated lesions can be reliably assessed when using this type of tool, while the presence of a ruler might need attention in the optimization of artificial intelligence tools for melanoma assessment.