Label-efficient medical image segmentation: the benefits and limitations of semi-supervised generative models
File(s)
Author(s)
Rosnati, Margherita
Type
Thesis
Abstract
Deep learning has displayed considerable promise in medical image segmentation for various applications, such as disease prognosis, quantifying disease progression, lesion volumes, tumour progression, and radiotherapy planning. However, the availability of ample annotated datasets for training is restricted and obtaining expert annotations is highly resource-intensive and time-consuming, making it troublesome to obtain a sizeable dataset. As a result, models are frequently trained on smaller datasets, which may not generalise well to changes in data characteristics, known as domain shifts.
To overcome the challenges of low data availability and domain shift, we focus on label-efficient paradigms and their robustness to environmental changes. The research presented in this thesis aims to achieve three primary objectives. Firstly, we intend to demonstrate the potential and pitfalls of machine learning in medical imaging through a case study in traumatic brain injury outcome prediction. We also emphasise the under-utilisation of data owing to annotation limitations. Secondly, we define an evaluation scheme for the robustness of Semi-Supervised Learning in medical image segmentation, focusing on data shifts and assessing the model's performance on downstream tasks such as classification. Our empirical analyses using a state-of-the-art generative model-based network highlight its generalisability and robustness in in- and out-of-domain tasks. Our evaluation framework contributes to the broader discussion on the efficacy of Semi-Supervised Learning methods in medical image segmentation. Finally, we propose a novel Semi-Supervised Learning method based on generative modelling that carefully utilises the latent features extracted from diffusion models, showcasing their ability to segment medical images robustly. Our experiments showcase the superiority of the proposed Timestep Ensembling Diffusion Model in diverse medical imaging tasks and illustrate the impact of mindful design choices.
To overcome the challenges of low data availability and domain shift, we focus on label-efficient paradigms and their robustness to environmental changes. The research presented in this thesis aims to achieve three primary objectives. Firstly, we intend to demonstrate the potential and pitfalls of machine learning in medical imaging through a case study in traumatic brain injury outcome prediction. We also emphasise the under-utilisation of data owing to annotation limitations. Secondly, we define an evaluation scheme for the robustness of Semi-Supervised Learning in medical image segmentation, focusing on data shifts and assessing the model's performance on downstream tasks such as classification. Our empirical analyses using a state-of-the-art generative model-based network highlight its generalisability and robustness in in- and out-of-domain tasks. Our evaluation framework contributes to the broader discussion on the efficacy of Semi-Supervised Learning methods in medical image segmentation. Finally, we propose a novel Semi-Supervised Learning method based on generative modelling that carefully utilises the latent features extracted from diffusion models, showcasing their ability to segment medical images robustly. Our experiments showcase the superiority of the proposed Timestep Ensembling Diffusion Model in diverse medical imaging tasks and illustrate the impact of mindful design choices.
Version
Open Access
Date Issued
2024-01
Date Awarded
2024-06
Copyright Statement
Creative Commons Attribution Licence
License URL
Advisor
Glocker, Benjamin
Sharp, David
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/S023283/1
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)