End-to-End Whole Slide Image Classification and Search using Set Representations

Abstract

Due to the sheer size of gigapixel whole slide images (WSIs), it is difficult to apply deep learning and computer vision algorithms for digital pathology. Traditional approaches rely on extracting meaningful patches from a WSI and obtaining a representation for each patch individually. This approach fails to incorporate inherent information between the set of extracted patches. In this thesis, we approach the problem of WSI representation by using Set Transformers, a neural network architecture capable of incorporating the element-wise interactions of sets to obtain one global representation. We show through extensive experiments the representation capabilities of our method by outperforming existing patch-based approaches on search and classification tasks.