Segmentation and Prediction of Mutation Status of Malignant Melanoma Whole-slide Images using Deep Learning

Abstract: Malignant melanoma is an aggressive type of skin cancer. Gene mutations can make the disease progress faster, but specialised treatment exists. Today, gene mutations are detected with DNA-analysis which is costly and time-consuming. The aim of our thesis is to investigate whether deep learning can be used to differentiate whole-slide images of tumours with different gene mutations. This was done in two steps, first whole-slide images were segmented based on tissue types, and then classification of gene mutations was done. The tissue segmentation was done using the deep convolutional network Inception v3, modified to a four class output. Image tiles of the size 244 x 244 pixels were used to train and evaluate the network, with F1-score 0.84 on tumour tissue. Two different methods to predict mutation status were tested. First, image features extracted from the segmentation network were fed into binary classifiers to separate images of tumours with and without NRAS mutation. Due to unsatisfactory results, another method was tested. A new Inception v3 network was trained to distinguish between NRAS and BRAF mutated tumours. Data from the public database The Cancer Genome Atlas was used for training and evaluation. Further testing was done on two independent test sets. Only tiles with 90% or higher probability of being tumour according to the segmentation network were used. The classification network was tested tilewise (AUC 0.53-0.66) and patientwise with AUC-values around 0.60 for all datasets. The results indicate that it is possible to separate tissue images based on gene mutations. We believe that deep learning networks like these have great potential of being integrated into diagnostics of malignant melanoma. This could lead to faster and more accessible gene mutation diagnostics around the world.