Building artificial intelligence algorithms to determine prognostic and predictive factors of melanoma from analysis of dermoscopic pictures.

Background: Melanoma is responsible for almost 90% of skin cancer deaths. Early diagnosis is crucial. Breslow thickness, the thickness of the tumor from granular layer to the deepest invasion, part of the “T” classification in TNM staging system, is an important prognostic factor. Currently, it is only measurable histologically, post-excision. Dermoscopy, a helpful tool for melanoma diagnosis, has limited power in evaluating melanoma thickness, (i.e. discriminating in-situ vs invasive melanoma). Among studies on artificial intelligence (AI) and melanoma, some focused on the evaluation of melanoma thickness, mainly with the task “in situ vs invasive melanoma”. Methods: The study investigated the use of AI to discriminate between in-situ and invasive melanoma based on dermoscopic images. 293 dermoscopic images (acquired with a Vidix 4.0 dermatoscope, with the standard software Dermagraphix of the manufacturer Medici Medical Srl) of confirmed melanomas, from 249 patients (116 female, 133 male), were collected. 161 images referred to in situ melanomas; 132 to invasive melanomas. Three different approaches were refined as data collection continued, thus each approach was tested on a gradually larger dataset. Nine dermatologists also participated in the study, evaluating the images in the dataset and classifying them in invasive melanomas vs in situ melanomas. Results: The experiment with the best result, involving a convolute neural network with fine-tuning, performed with an overall accuracy of 63.95%. Namely, 60/132 (45.45%) invasive melanomas and 132/161 (81.99%) in situ melanomas were correctly recognized. Nine dermatologists with expertise in dermoscopy achieved an overall accuracy of 61.29%. Namely, they correctly identified 64.74% of invasive melanomas. In situ melanomas were properly recognized with a class accuracy of 58.45%. Discussion: Although probably not enough to be integrated in the clinical routine, an overall accuracy of 63.95% is aligned with the literature, despite a considerably smaller sample size. Moreover, dermatologists consistently outperformed AI tools in all these studies. Oppositely, in the present study, CNNs outperformed dermatologists in terms of overall accuracy. In terms of class-specific accuracy, CNNs performed better in recognizing in situ melanomas (81.99% vs 58.45%), whereas dermatologists outperformed CNNs in identifying invasive melanomas (45.45% vs 64.74%). This highlights the importance of class-specific performance in clinical contexts. The main limitation of this preliminary study is the small sample size, which hinders the use of state-of-the-art AI methods and makes the algorithms more susceptible to overfitting. Also, the vast majority (86.36%) of invasive melanomas was represented by pT1a melanomas. This results in two similar classes, making the classification task particularly difficult, but also adherent to the “real life” daily challenges. This challenging case is overlooked in the literature, where the focus is on melanomas with more distinct characteristics, making classification easier. Despite that, the performance of this study still achieves state-of-the-art results. Though the current accuracy is limited and not acceptable for deployment in the clinical routine, these findings suggest that by continuing to expand the sample, and employing increasingly sophisticated methods, there is room to significantly improve these results.