Discrimination Between Invasive and In Situ Melanomas Using Clinical Close-Up Images and a De Novo Convolutional Neural Network

• Main Authors: Sam Polesie, Martin Gillstedt, Gustav Ahlgren, Hannah Ceder, Johan Dahlén Gyllencreutz, Julia Fougelberg, Eva Johansson Backman, Jenna Pakka, Oscar Zaar, John Paoli
• Format: Article
• Language: English
• Published: Frontiers Media S.A. 2021-09-01
• Series: Frontiers in Medicine
• Subjects: artificial intelligence; clinical decision-making; melanoma; neural networks; computer; supervised machine learning
• Online Access: https://www.frontiersin.org/articles/10.3389/fmed.2021.723914/full

Background: Melanomas are often easy to recognize clinically but determining whether a melanoma is in situ (MIS) or invasive is often more challenging even with the aid of dermoscopy. Recently, convolutional neural networks (CNNs) have made significant and rapid advances within dermatology image analysis. The aims of this investigation were to create a de novo CNN for differentiating between MIS and invasive melanomas based on clinical close-up images and to compare its performance on a test set to seven dermatologists.Methods: A retrospective study including clinical images of MIS and invasive melanomas obtained from our department during a five-year time period (2016–2020) was conducted. Overall, 1,551 images [819 MIS (52.8%) and 732 invasive melanomas (47.2%)] were available. The images were randomized into three groups: training set (n = 1,051), validation set (n = 200), and test set (n = 300). A de novo CNN model with seven convolutional layers and a single dense layer was developed.Results: The area under the curve was 0.72 for the CNN (95% CI 0.66–0.78) and 0.81 for dermatologists (95% CI 0.76–0.86) (P < 0.001). The CNN correctly classified 208 out of 300 lesions (69.3%) whereas the corresponding number for dermatologists was 216 (72.0%). When comparing the CNN performance to each individual reader, three dermatologists significantly outperformed the CNN.Conclusions: For this classification problem, the CNN was outperformed by the dermatologist. However, since the algorithm was only trained and validated on 1,251 images, future refinement and development could make it useful for dermatologists in a real-world setting.