Basal cell carcinomas (BCCs) represent one of the most common cutaneous malignancies in humans. Because they are so abundant, pathologists have little time to diagnose a single lesion. The aim of the study was to test whether it is possible to detect basal cell carcinomas in anonymized histologic sections using state-of-the-art machine learning methods. Moreover, we aimed to identify the key image regions that are important for the decision of the artificial intelligence.
We implemented a convolutional neural network (CNN) that was trained to classify histological H&E sections as ‘BCC’ or ‘no BCC’. Images were classified using multiple instance learning and attention. Standard statistical parameters like AUC and accuracy were calculated. For retrieving ‘diagnostic key regions’ of the images, we used integrated gradients and attention methods. Those were compared to ‘diagnostic key regions’ for pathologists, which were determined using eye-tracking techniques.
The CNN was trained on 709 training images. When testing it on 129 previously unseen images, it decided correctly in 95 out of 100 cases and received an AUC of 0.99. With the help of the ‘attention’ method we were able to trace the most important ‘key’ regions for the classification of the neural network. As compared to four board-proven pathologists the CNN used partly different, new ‘diagnostic key regions’ for its decision.

Neural networks detect cutaneous basal cell carcinomas in histological sections