AIS-CaP · Interpretable Artificial Intelligence across Scales for Next-Generation Cancer Prognostics

Computation pathology has the potential to revolutionize cancer care and research, specifically through improving assessment of patient prognosis and treatment selection by applying advanced machine learning methods to digitized tissue sections, i.e. whole-slide images (WSIs). This will allow us to replace the current state-of-the-art of human-developed cancer grading systems. However, the field is currently hindered by significant knowledge gaps: we do not know how to effectively leverage both global and local information in WSIs, how to identify pan-cancer prognostic features, and how to make machine learning models explainable and interpretable. In this project, I will address these key knowledge gaps by building on the novel stochastic streaming gradient descent developed in my group. Specifically, I will integrate innovative multi-task and cross-task learning algorithms with SSGD. Furthermore, I will leverage the latest advances in self-supervision, self-attention and natural language processing to endow deep neural networks with unprecedented transparency and explainability. Last, the project will validate our developed methodology in the largest dataset of oncological WSIs in the world, and, for the first time, identify links between morphological prognostic features and genetic features. By publicly releasing all developed tools and data, the proposed project will have a scientific multiplier effect for the fields of oncology, computational pathology and machine learning. Specifically, the derived cancer-specific and pan-cancer biomarkers can be leveraged in clinical care and cancer research, the enhanced SSGD method for other tasks in computational pathology and our novel multi-task and explainability algorithms can impact other research areas in machine learning, such as remote sensing and self-driving cars.