Bieder, Florentin. Memory-efficient deep learning methods for brain image analysis. 2024, Doctoral Thesis, University of Basel, Faculty of Medicine.
![]()
|
PDF
7Mb |
Official URL: https://edoc.unibas.ch/96841/
Downloads: Statistics Overview
Abstract
The diverse array of imaging modalities currently in use has profoundly impacted the practice of medicine. In particular, X-ray computed tomography (CT) and magnetic resonance imaging (MRI) have the capacity to record a three-dimensional image of the body with great detail. As a consequence of their widespread use, the amount of available data is continuously increasing at an unprecedented rate. This presents an opportunity to apply machine learning and, in particular, deep learning. Deep learning is a subfield of machine learning that is particularly well-suited to train on large quantities of data and to capture complex features in high-dimensional data.
Currently, deep learning methods rely on graphics processing units (GPUs) to be evaluated with any reasonable speed. However, GPUs with high-performance characteristics – which are directly linked to their memory capacity – contribute significantly to the high implementation costs of deep learning methods. To enable an easier access and a more widespread adoption, it is crucial to reduce this barrier of entry.
In this work, we explore the potential of deep learning methods for a range of tasks involving brain imaging data, with a focus on reducing the GPU-memory consumption. This is particularly relevant for methods that process MR and CT scans, as they inherently are three-dimensional. The first project we present is an unsupervised anomaly detection and localization method for brain CT scans. It employs a novel self-supervised surrogate task. The second project involves a supervised tumor segmentation in MR scans with denoising diffusion models. This method has previously been proposed for two-dimensional slices. As diffusion models already require substantial resources for two- dimensional data, we explore various techniques to reduce the resource consumption, to adapt this model to three-dimensional data. In the last project, we explore implicit neural representations (INRs) to model the development of the neonatal brain on the basis of MR scans. We show that INRs can be trained in sparsely sampled data, explore techniques for disentangling the latent space and illustrate how they can be trained with minimal resources.
Currently, deep learning methods rely on graphics processing units (GPUs) to be evaluated with any reasonable speed. However, GPUs with high-performance characteristics – which are directly linked to their memory capacity – contribute significantly to the high implementation costs of deep learning methods. To enable an easier access and a more widespread adoption, it is crucial to reduce this barrier of entry.
In this work, we explore the potential of deep learning methods for a range of tasks involving brain imaging data, with a focus on reducing the GPU-memory consumption. This is particularly relevant for methods that process MR and CT scans, as they inherently are three-dimensional. The first project we present is an unsupervised anomaly detection and localization method for brain CT scans. It employs a novel self-supervised surrogate task. The second project involves a supervised tumor segmentation in MR scans with denoising diffusion models. This method has previously been proposed for two-dimensional slices. As diffusion models already require substantial resources for two- dimensional data, we explore various techniques to reduce the resource consumption, to adapt this model to three-dimensional data. In the last project, we explore implicit neural representations (INRs) to model the development of the neonatal brain on the basis of MR scans. We show that INRs can be trained in sparsely sampled data, explore techniques for disentangling the latent space and illustrate how they can be trained with minimal resources.
Advisors: | Cattin, Philippe Claude |
---|---|
Committee Members: | Psychogios, Marios-Nikos and Heinrich, Mattias Paul and Sankühler, Robin and Pezold, Simon |
Faculties and Departments: | 03 Faculty of Medicine > Departement Biomedical Engineering > Imaging and Computational Modelling > Center for medical Image Analysis & Navigation (Cattin) |
UniBasel Contributors: | Cattin, Philippe Claude and Psychogios, Marios-Nikos and Pezold, Simon |
Item Type: | Thesis |
Thesis Subtype: | Doctoral Thesis |
Thesis no: | 15606 |
Thesis status: | Complete |
Number of Pages: | xii, 112 |
Language: | English |
Identification Number: |
|
edoc DOI: | |
Last Modified: | 31 Jan 2025 05:30 |
Deposited On: | 30 Jan 2025 11:52 |
Repository Staff Only: item control page