Performance of a Deep Learning Algorithm for Melanoma Classification Across Diverse Dermoscopic and Tele‐Dermatology Datasets

ABSTRACT Background Early detection of melanoma significantly boosts patient survival rates. Deep learning has demonstrated dermatologist‐level accuracy in assessing pigmented skin lesions by analysing images at the pixel level. However, these neural networks may face challenges with ‘real‐life’ images due to limited training data and image artefacts. Objectives Our study aimed to create an innovative deep learning algorithm and further assess its performance in classifying melanoma and non‐melanoma lesions across a range of datasets, beginning with publicly available dermoscopic image databases, followed by dermoscopic images collected by dermatologists, and concluding with real‐life pictures taken by primary care professionals for tele‐dermatology. Methods We trained our model with the ISIC large data set of 65,109 unique labelled images after filtering. An innovative optimization strategy was applied to create a ‘doubtful’ category derived from the two original categories: melanoma and non‐melanoma. We then assessed the performance of our algorithm on the GLOMEL database of 2672 dermoscopic images in comparison with pathological results, and on 294 images of pigmented lesions retrospectively collected from the outpatient dermatology clinics, and two tele‐medicine platforms, in comparison with experts' recommendations. Results Our algorithm demonstrated a sharp increase of performance when communicating on ‘doubtful’ cases versus the traditional binary approach. With this new method, it achieved a robust performance in comparison to dermatologists in melanoma classification using a variety of dermoscopy images and high‐resolution clinical images. This method is clinically relevant because it highlights, via the ‘doubtful’ category, which lesions require the insight of human experts and which ones are clear‐cut cases. Conclusions These results demonstrate HUVY's potential to help primary care providers make more informed referrals by distinguishing between likely melanomas, “doubtful” lesions and clearly benign lesions. Freed from ⅔ of demands on benign lesions it allows specialists to focus on these most critical or uncertain cases.