Evaluation of Melanoma Thickness with Clinical Close-up and Dermoscopic Images Using a Convolutional Neural Network

Abstract

Convolutional neural networks (CNNs) have shown promise in discriminating between invasive and in situ melanomas. The aim of this study was to analyse how a CNN model, integrating both clinical close-up and dermoscopic images, performed compared with 6 independent dermatologists. The secondary aim was to address which clinical and dermoscopic features dermatologists found to be suggestive of invasive and in situ melanomas, respectively. A retrospective investigation was conducted including 1,578 cases of paired images of invasive (n = 728, 46.1%) and in situ melanomas (n = 850, 53.9%). All images were obtained from the Department of Dermatology and Venereology at Sahlgrenska University Hospital and were randomized to a training set (n = 1,078), a validation set (n = 200) and a test set (n = 300). The area under the receiver operating characteristics curve (AUC) among the dermatologists ranged from 0.75 (95% confidence interval 0.70–0.81) to 0.80 (95% confidence interval 0.75–0.85). The combined dermatologists’ AUC was 0.80 (95% confidence interval 0.77–0.86), which was significantly higher than the CNN model (0.73, 95% confidence interval 0.67–0.78, p = 0.001). Three of the dermatologists significantly outperformed the CNN. Shiny white lines, atypical blue-white structures and polymorphous vessels displayed a moderate interobserver agreement, and these features also correlated with invasive melanoma. Prospective trials are needed to address the clinical usefulness of CNN models in this setting.

Keywords

• artificial intelligence
• clinical decision-making
• melanoma
• neural networks, computer
• supervised machine learning