Fully Convolutional Network for Melanoma Diagnostics

This work seeks to determine how modern machine learning techniques may be applied to the previously unexplored topic of melanoma diagnostics using digital pathology. We curated a new dataset of 50 patient cases of cutaneous melanoma using digital pathology. We provide gold standard annotations for three tissue types (tumour, epidermis, and dermis) which are important for the prognostic measurements known as Breslow thickness and Clark level. Then, we devised a novel multi-stride fully convolutional network (FCN) architecture that outperformed other networks trained and evaluated using the same data according to standard metrics. Finally, we trained a model to detect and localize the target tissue types. When processing previously unseen cases, our model's output is qualitatively very similar to the gold standard. In addition to the standard metrics computed as a baseline for our approach, we asked three additional pathologists to measure the Breslow thickness on the network's output. Their responses were diagnostically equivalent to the ground truth measurements, and when removing cases where a measurement was not appropriate, inter-rater reliability (IRR) between the four pathologists was 75.0%. Given the qualitative and quantitative results, it is possible to overcome the discriminative challenges of the skin and tumour anatomy for segmentation using modern machine learning techniques, though more work is required to improve the network's performance on dermis segmentation. Further, we show that it is possible to achieve a level of accuracy required to manually perform the Breslow thickness measurement.