Learning diffusion MR from commercially available protocols
: bringing advanced tractography into routine neurosurgical practice

Student thesis: Doctoral ThesisDoctor of Philosophy


Preserving eloquent white matter (WM) regions during brain surgery is key to reducing post-operative deficits, and consequently preserving patient quality of life. Tractography can be used to map WM tracts pre-operatively in a non-invasive manner by estimating the directional distribution of myelinated fibers from measured water diffusivity of brain tissues on diffusion magnetic resonance imaging (dMRI). To obtain high-quality local fiber orientation distributions (FODs), which are necessary for reliable tractography, it is required acquisition parameters with high signal magnitude (b-values) and a high number of diffusion-weighting gradients (b-vectors). However, in clinical settings scanner time is limited, and many clinical scanners may not provide appropriate state-of-the-art dMRI acquisitions necessary for high-quality tractography. Furthermore, validating tractography in a clinical setting is an active research topic as gold standard ground truth for the location of WM tracts is not possible to obtain for clinical data. Tractography is unreliable for patients with tumors or lesions because the tracts are stretched or shifted with respect to normal anatomy, which hinders the use of tractography in a clinical setting.

Convolutional neural networks (CNNs) have emerged as the state-of-the-art approach for most image segmentation, regression, and classification tasks related to brain structures. These networks use annotated datasets to learn underlying mathematical representations, which enable them to accurately perform specific tasks. In this thesis, I present novel CNN-based methods that advance the use of tractography and tract segmentation in dMRI data acquired from clinically available protocols.

Firstly, I propose a method that utilizes 3D CNNs to generate high-quality FODs from single-shell dMRI. Single-shell dMRI is the most commonly used dMRI acquisition in clinical settings. Secondly, I present a reliable tract segmentation method that incorporates uncertainty quantification. This method utilizes a 3D UNet to segment white matter tracts and estimate uncertainty in the model and data using test-time dropout and test-time augmentation. A volume-based calibration approach is leveraged to compute representative predicted probabilities from the estimated uncertainties.

Thirdly, I investigate the correlation between the uncertainty output from our 3D U-Net tract segmentation algorithm and transcranial magnetic stimulation of the motor cortex in patients with tumors. I demonstrate that our tract segmentation method generalizes to datasets acquired clinically on patients with brain pathologies that distort normal anatomy. Throughout all these methods, I compare results with ground truth datasets, qualitative assessments, and correlation measurements.

The findings of this thesis demonstrate that 3D CNNs can accurately enhance FOD models from dMRI, segment tract locations, and estimate tract uncertainty. These DL-based solutions can improve tract segmentations and enable clinicians to make more informed choices during pre-operative surgical planning and safety assessment. The significance of this thesis is that it provides DL-based solutions to augment dMRI and improve tract segmentations to enable clinicians to make more informed choices.

Date of Award1 Nov 2023
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorSebastien Ourselin (Supervisor) & Rachel Sparks (Supervisor)

Cite this