Goodwin-Allcock, Tobias; (2024) Deep learning enabled enhancement of clinical diffusion tensor imaging for stroke. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Goodwin-Allcock_10196353_Thesis.pdf Download (65MB) | Preview

Abstract

This thesis develops machine learning methods for obtaining diffusion tensor (DT) parameters from the sparse data acquisitions clinical scanner demands compel. Clinical management may be aided by DT characterisation of brain damage. However, accurate DT estimation with conventional model fitting (MF) techniques requires more measurements than are clinically tolerable, a problem theoretically addressable with machine learning (ML)-estimated DT estimation. The application of existing ML methods is inhibited by their large data requirements and the marked pathological, biological, and instrumental variability characteristics of the clinical context. Clinically compatible ML has not yet obtained directional information. The first contribution of this thesis locates the cause of certain ML model failures to ignorance of the geometry of diffusion data. I show that accounting for geometry in the ML method increases robustness while decreasing training data requirements. The second contribution develops a new ML method, Patch-CNN, able to estimate directional information from sparse, clinical-grade imaging with high fidelity. Patch-CNN employs a convolutional window to estimate directional parameters from local neighbourhood information. Window size is kept small to minimize training data demands. I show that Patch-CNN can reveal major anatomical structures, such as the corticospinal tract, in clinical scans with only a single training volume. The third and fourth contributions tests the robustness of Patch-CNN to pathology unseen during training. The model is trained on data free of stroke lesions and tested on data containing them. In the third contribution, highly sampled imaging of people who all had a stroke over 5 days before scanning are used to analyse Patch-CNN's robustness. Model fidelity is shown to be invariant to the presence of lesions, permitting reliable estimation of both normal and abnormal microstructural anatomy even without training on pathological data. Finally we apply Patch-CNN to clinically acquired stroke data where there is no high quality ground truth diffusion parameter values. Patch-CNN estimates diffusion tensor parameters with higher signal to noise than traditional model fitting.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item