Machine learning for brain lesion segmentation
File(s)
Author(s)
Basaran, Berke Doga
Type
Thesis
Abstract
Deep learning techniques hold significant promise in advancing medical image processing. This is especially true for brain lesion segmentation, which is crucial for diagnosing and monitoring neurological conditions like multiple sclerosis (MS). However, challenges such as lack of clinically relevant machine learning models, data scarcity, and model generalisation remain major obstacles. This thesis addresses these challenges by proposing innovative machine learning methodologies for improving brain lesion segmentation.
The first contribution of this work focuses on a clinically overlooked task: segmentation of new lesions in MS using longitudinal data. We develop an end-to-end pipeline leveraging a convolutional neural network with specialised preprocessing and data augmentation techniques to segment new lesions appearing in follow-up scans. We validate our method against a public challenge dataset, and demonstrate superior performance against state-of-the-art biomedical segmentation models.
The second focus is on overcoming the scarcity of labelled data, which hampers the training of effective segmentation models. Two approaches are proposed. The first utilises an adversarial technique to generate synthetic brain images with lesions from healthy subjects, enhancing training dataset diversity. The second approach augments datasets at the lesion-level by altering the lesion load of images with respect to dataset load distributions. Both methods enhance brain lesion segmentation performance, especially in data-limited scenarios.
Lastly, this thesis tackles the challenge of model generalisation by introducing a segmentation model capable of handling heterogeneous input data and segmenting multiple lesion types. Our model incorporates domain-specific anatomical constraints and benefits from the augmented datasets produced by LesionMix. Results show that allowing for diverse input data and incorporating domain knowledge, we can achieve enhanced segmentation accuracy and generalisability across different datasets.
Collectively, these contributions provide a robust framework for advancing brain lesion segmentation, with significant implications for clinical applications and the broader field of medical image analysis.
The first contribution of this work focuses on a clinically overlooked task: segmentation of new lesions in MS using longitudinal data. We develop an end-to-end pipeline leveraging a convolutional neural network with specialised preprocessing and data augmentation techniques to segment new lesions appearing in follow-up scans. We validate our method against a public challenge dataset, and demonstrate superior performance against state-of-the-art biomedical segmentation models.
The second focus is on overcoming the scarcity of labelled data, which hampers the training of effective segmentation models. Two approaches are proposed. The first utilises an adversarial technique to generate synthetic brain images with lesions from healthy subjects, enhancing training dataset diversity. The second approach augments datasets at the lesion-level by altering the lesion load of images with respect to dataset load distributions. Both methods enhance brain lesion segmentation performance, especially in data-limited scenarios.
Lastly, this thesis tackles the challenge of model generalisation by introducing a segmentation model capable of handling heterogeneous input data and segmenting multiple lesion types. Our model incorporates domain-specific anatomical constraints and benefits from the augmented datasets produced by LesionMix. Results show that allowing for diverse input data and incorporating domain knowledge, we can achieve enhanced segmentation accuracy and generalisability across different datasets.
Collectively, these contributions provide a robust framework for advancing brain lesion segmentation, with significant implications for clinical applications and the broader field of medical image analysis.
Version
Open Access
Date Issued
2024-09-14
Date Awarded
01/02/2025
License URL
Advisor
Bai, Wenjia
Matthews, Paul
Sponsor
UK Research and Innovation
Grant Number
EP/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)