Deep learning for histology-based cancer research and diagnostics

Nearly 20 million people around the world were diagnosed with cancer in 2020 causing almost 10 million lives to be lost. Female breast cancer has surpassed lung cancer as the most frequent type with 2.3 million women diagnosed with breast cancer every year. The total global cancer burden is expected to rise by 47% from 2020 to be 28.4 million cases in 2040. The machine learning community has an obligation to use the power of deep neural networks to look for novel and sustainable solutions that make a true impact on the field of pathology - the rock bed of cancer research and diagnostics. Currently, the leaps in development are often confined to the proof-of-concept stage, never reaching the end-users. The main challenges are related to both complex diagnostic regimes, lack of standardization, and the cost of obtaining training data. These aspects make it difficult to build algorithms that generalize into the clinical domain. The goal of this thesis is to investigate some of the challenges of bringing algorithms into real-world settings in pathology. By studying realistic shifts in data distributions, we show that deep neural networks can generalize to and provide reliable uncertainty estimates within the cancer indication it was trained on. On the contrary, they fail to report rare abnormalities, and other systems need to inspect incoming data for signs of significant changes in input distributions. Moreover, we demonstrate that it is possible to create an automatic computational approach to quantify the tumor infiltrating lymphocytes (TILs) that could help in standardizing the prognostic assessment. In turn, this can support clinicians in treatment decisions to provide better patient outcomes. Meanwhile, we also investigate a more objective and scalable method to create training labels. Finally, we assess some of the remaining challenges of using machine learning, and how these can be overcome with further research.