Spatial context in computational pathology
Muhammad Shaban · Warwick Research Archive Portal (University of Warwick) · 2020
In recent years, computational pathology has emerged as a discipline representing big-data based approaches for the diagnosis and prognosis of cancer patients using different sources of data, mainly digitised histology images and clinical information. A plethora of computational methods have been developed for fast and reproducible diagnosis and prognosis of cancer, lately dominated by deep learning based methods. However, current deep learning methods do not incorporate the whole spatial landscape of histology images due to limited computational and memory resources. In this thesis, I develop deep learning based methods which incorporate the broader spatial context of histology images for cancer diagnosis and prognosis. I propose a novel framework to incorporate large contextual information inheritably available in histology images by a context-aware neural network. The proposed framework first encodes the local representation of an input image into low dimensional features then aggregates the features by considering their spatial organization to make a final prediction. The framework is designed for a set of histology problems which requires both high-resolution appearance of tissue along with large contextual information such as colorectal grading, and growth pattern classification. I have also proposed two novel objective measures for the quantification of tumour microenvironment of head and neck squamous cell carcinoma (HNSCC) patients for their better stratification and prognostication. The first measure quantifies the tumour infiltrating lymphocytes abundance (TILAb) whereas the second one is for the quantification of tumour-associated stroma infiltrating lymphocytes to tumour-associated stroma ratio (TASIL-Ratio). Both TILAb and TASIL-Ratio based scores show prognostic significance similar to manual scores but with the added advantages of a more rapid and objective quantification.