Recognition of Cutaneous Melanoma on Digitized Histopathological Slides via Artificial Intelligence Algorithm

• Main Authors: Francesco De Logu, Filippo Ugolini, Vincenza Maio, Sara Simi, Antonio Cossu, Daniela Massi, Italian Association for Cancer Research (AIRC) Study Group, Romina Nassini, Marco Laurino
• Format: Article
• Language: English
• Published: Frontiers Media S.A. 2020-08-01
• Series: Frontiers in Oncology
• Subjects: cutaneous melanoma; artificial intelligence; convolutional neural network; image analysis; diagnosis
• Online Access: https://www.frontiersin.org/article/10.3389/fonc.2020.01559/full

Increasing incidence of skin cancer combined with a shortage of dermatopathologists has increased the workload of pathology departments worldwide. In addition, the high intraobserver and interobserver variability in the assessment of melanocytic skin lesions can result in underestimated or overestimated diagnosis of melanoma. Thus, the development of new techniques for skin tumor diagnosis is essential to assist pathologists to standardize diagnoses and plan accurate patient treatment. Here, we describe the development of an artificial intelligence (AI) system that recognizes cutaneous melanoma from histopathological digitalized slides with clinically acceptable accuracy. Whole-slide digital images from 100 formalin-fixed paraffin-embedded primary cutaneous melanoma were used to train a convolutional neural network (CNN) based on a pretrained Inception-ResNet-v2 to accurately and automatically differentiate tumoral areas from healthy tissue. The CNN was trained by using 60 digital slides in which regions of interest (ROIs) of tumoral and healthy tissue were extracted by experienced dermatopathologists, while the other 40 slides were used as test datasets. A total of 1377 patches of healthy tissue and 2141 patches of melanoma were assessed in the training/validation set, while 791 patches of healthy tissue and 1122 patches of pathological tissue were evaluated in the test dataset. Considering the classification by expert dermatopathologists as reference, the trained deep net showed high accuracy (96.5%), sensitivity (95.7%), specificity (97.7%), F1 score (96.5%), and a Cohen’s kappa of 0.929. Our data show that a deep learning system can be trained to recognize melanoma samples, achieving accuracies comparable to experienced dermatopathologists. Such an approach can offer a valuable aid in improving diagnostic efficiency when expert consultation is not available, as well as reducing interobserver variability. Further studies in larger data sets are necessary to verify whether the deep learning algorithm allows subclassification of different melanoma subtypes.