DEEP MULTIPLE INSTANCE LEARNING ON BIOLOGICAL DATA

Abstract

This thesis provides a study into the implementation of Deep Multiple Instance Learning (MIL) models to problems in biological data, specifically the detection of m6A RNA modification from direct RNA sequencing data and classification of diabetic retinopathy from funduscopic images. The two problem domains demonstrate a structure that MIL can effectively solve and we are going to show in this thesis that the application of Deep MIL models to these two problems solves practical needs that arise naturally from the deployment of machine learning models in the two domains. Our implementation of Deep MIL models solves not only the basic classification problems in the two domains but also the identification of key instances for stoichiometry prediction in RNA modifications as well as the detection of ROIs in diabetic retinopathy classification.