Automatic detection of histopathological melanoma images using an Artificial Intelligence algorithm

Background Melanoma is among the most commonly diagnosed skin cancers worldwide and its recognition requires visual inspection by an experienced pathologist. To assist pathologists in accurate interpretation, automatic analysis on melanoma whole-slide images (WSI) has been studied to predict diagnosis and classification. Recently, the combination of image analysis and optical microscopy has been remarkably improved by the development of neural networks and dedicated algorithms. In this study, we implemented an annotation framework for automated image analysis which makes use of an Artificial Intelligence (AI) algorithm that allows masks to be built on the scanned tissues to perform targeted analysis for melanoma recognition. Methods Anonymized histopathological slides stained with haematoxylin and eosin representative of 100 cutaneous primary melanomas >2 mm in Breslow thickness (62M; 38F; mean age 63 yrs, range 24-90 yrs; pT Stage: pT3=48, pT4=58; 71 with ulceration and 29 w/o ulceration) were selected for the study and digitalized using Aperio Digital Pathology (Leica Biosystems) with X40 power.From each scanned slide, an expert dermatopathologist extracted Region-Of-Interests (ROIs) related to melanoma and to normal tissue, respectively. Each ROI was cropped in non-overlapping 299×299 pixel tiles (20-fold magnification). The tiles obtained from ROIs of 60 slides (normal tissue: 1377 tiles, melanoma: 2141 tiles) were used as training-dataset, the tiles from ROI of remaining 40 slides (normal tissue: 791 tiles, melanoma: 1122 tiles) as test-dataset. We trained a deep convolutional neural network (based on a pre-trained Inception-ResNet-v2) with the training-dataset, by using 75% of the tiles for training and 25% for validation. We freezed the first 10 layers of the net, and set the learning rate to 0.003 for the residual layers. Subsequently, we tested the net performance with the test-dataset. Results Training using the curated image patches took approximately 18h to complete 3200 iterations with Matlab software (R2019b, Deep Learning Toolbox). The overall accuracy of our net was 96.5% (1847 correct classification on a total of 1913). Specifically, the misclassification rates were 2.3% for normal tissue (18 tiles on 791) and 4.3% for melanoma (48 tiles on 1122). Conclusions Our findings indicate that AI algorithms can obtain high performance in the evaluation of histopathological melanoma images. Prospective studies implemented in a clinical setting are necessary to confirm the potential clinical utility of AI in assisting pathologists in cutaneous melanoma diagnoses.