Collaborative learning of joint medical image segmentation tasks from heterogeneous and weakly-annotated data

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

Convolutional Neural Networks (CNNs) have become the state-of-the-art for most image segmentation tasks and therefore one would expect them to be able to learn joint tasks, such as brain structures and pathology segmentation. However, annotated databases required to train CNNs are usually dedicated to a single task, leading to partial annotations (e.g. brain structure or pathology delineation but not both for joint tasks). Moreover, the information required for these tasks may come from distinct magnetic resonance (MR) sequences to emphasise different types of tissue contrast, leading to datasets with different sets of image modalities. Similarly, the scans may have been acquired at different centres, with different MR parameters, leading to differences in resolution and visual appearance among databases (domain shift). Given the large amount of resources, time and expertise required to carefully annotate medical images, it is unlikely that large and fully-annotated databases will become readily available for every joint problem. For this reason, there is a need to develop collaborative approaches that exploit existing heterogeneous and task-specific datasets, as well as weak annotations instead of time-consuming pixel-wise annotations.

In this thesis, I present methods to learn joint medical segmentation tasks from task-specific, domain-shifted, hetero-modal and weakly-annotated datasets. The problem lies at the intersection of several branches of Machine Learning: Multi-Task Learning, Hetero-Modal Learning, Domain Adaptation and Weakly Supervised Learning. First, I introduce a mathematical formulation of a joint segmentation problem under the constraint of missing modalities and partial annotations, in which Domain Adaptation techniques can be directly integrated, and a procedure to optimise it. Secondly, I propose a principled approach to handle missing modalities based on Hetero-Modal Variational Auto-Encoders. Thirdly, in this thesis, I focus on Weakly Supervised Learning techniques and present a novel approach to train deep image segmentation networks using particularly weak train-time annotations: only 4 (2D) or 6 (3D) extreme clicks at the boundary of the objects of interest. The proposed framework connects the extreme points using a new formulation of geodesics that integrates the network outputs and uses the generated paths for supervision. Fourthly, I introduce a new weakly-supervised Domain Adaptation technique using scribbles on the target domain and formulate as a cross-domain CRF optimisation problem. Finally, I led the organisation of the first medical segmentation challenge for unsupervised cross-modality domain adaptation (crossMoDA). The benchmark reported in this thesis provides a comprehensive characterisation of cross-modality domain adaptation techniques.

Experiments are performed on brain MR images from patients with different types of brain diseases: gliomas, white matter lesions and vestibular schwannoma. The results demonstrate the broad applicability of the presented frameworks to learn joint segmentation tasks with the potential to improve brain disease diagnosis and patient management in clinical practice.
Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorTom Vercauteren (Supervisor) & Sebastien Ourselin (Supervisor)

Cite this

'