Predicting melanoma patient outcomes using digital pathology

Melanoma is the most aggressive form of skin cancer and fifth most common cancer in the UK. Although immunotherapy has improved outcomes for patients with advanced stages of the disease, some patients do not respond to these treatments or develop resistance during therapy. Additionally, the current staging system which is used for determining prognosis and guiding the treatment of melanoma patients, shows significant variability, with some early-stage patients progressing to metastatic disease. Therefore, identifying which patients will respond to treatments and discovering new prognostic biomarkers have become crucial research areas for improving patient outcomes. The development of digital slide scanners has meant that tissue slides, which contain a wealth of phenotypic information, are being digitised to generate whole slide images (WSIs). This thesis investigates how artificial neural networks with digital pathology workflows, can be used to stratify patients based on their outcomes.

Firstly, we classified patients' WSIs into subgroups based on the inferred expression of immune cells within their tumors. These immune subgroups, derived from genetic data, present varying potential treatment targets and survival outcomes. WSIs are multi-resolution, multi-gigabyte images containing billions of pixels but often will only have a slide-level label. To handle this, we employed multiple instance learning (MIL) methods, where the image is divided into numerous patches, and the image is classified based on a set of these patches using a single supervisory label. We are among the first to investigate how factors such as patch resolution, feature extraction methods, and MIL techniques influence the classification of melanoma patients into immune subgroups. In a primary melanoma dataset, we achieve a mean area under the receiver operating characteristic curve (AUC) of 0.80 for classifying histopathology images into high or low immune subgroups and a mean AUC of 0.82 in an independent TCGA melanoma dataset. Our findings indicate that using 10x resolution patches, pathology-specific feature extraction methods, and attention-based MIL models improve classification performance.

To build on these MIL methods, we introduce a novel way to represent WSIs, using multi-resolution patch graphs, with resolution-aware node embeddings. These graphs enable us to capture long-range dependencies within the images. By employing graph neural networks to aggregate surrounding node embeddings from the WSI patches, we improved the classification performance beyond the MIL models. Here, we achieved a mean test AUC of 0.81 for classifying low and high immune melanoma subtypes, using graph representations.

Finally, we use survival graph neural networks to identify new prognostic subgroups from patch graph WSI representations. Here we show that the risk groups and immune subgroups generated from the models presented within this thesis were predictive of melanoma-specific survival (Concordance index = 0.73), even when adjusting for known prognostic factors. This suggests that these risk groups could represent novel predictive and prognostic biomarkers. Overall this thesis adds to the evidence that digital pathology workflows are an emerging tool for better understanding melanoma patient outcomes and survival.