Machine Learning Approaches for Semantic Segmentation on Partly-Annotated Medical Images

Abstract

Semantic segmentation of medical images plays a crucial role in assisting medical practitioners in providing accurate and swift diagnoses; nevertheless, deep neural networks require extensive labelled data to learn and generalise appropriately. This is a major issue in medical imagery because most of the datasets are not fully annotated. Training models with partly-annotated datasets generate plenty of predictions that belong to correct unannotated areas that are categorised as false positives; as a result, standard segmentation metrics and objective functions do not work correctly, affecting the overall performance of the models. In this thesis, the semantic segmentation of partly-annotated medical datasets is extensively and thoroughly studied. The general objective is to improve the segmentation results of medical images via innovative supervised and semi-supervised approaches. The main contributions of this work are the following. Firstly, a new metric, specifically designed for this kind of dataset, can provide a reliable score to partly-annotated datasets with positive expert feedback in their generated predictions by exploiting all the confusion matrix values except the false positives. Secondly, an innovative approach to generating better pseudo-labels when applying co-training with the disagreement selection strategy. This method expands the pixels in disagreement utilising the combined predictions as a guide. Thirdly, original attention mechanisms based on disagreement are designed for two cases: intra-model and inter-model. These attention modules leverage the disagreement between layers (from the same or different model instances) to enhance the overall learning process and generalisation of the models. Lastly, innovative deep supervision methods improve the segmentation results by training neural networks one subnetwork at a time following the order of the supervision branches. The methods are thoroughly evaluated on several histopathological datasets showing significant improvements.