Automatic white matter bundle segmentation in glioma patients

In glioma patients, the virtual dissection of white matter tracts is essential for surgical planning, as it supports the resection of the tumor while minimizing the risk of postoperative neurological impairments. While the automation of this task through data-driven approaches has been extensively explored in the literature for healthy individuals-benefiting from the availability of large datasets and the consistency of the white matter anatomy aligned with canonical references—its application to glioma patients presents two major challenges: (i) large, openly available clinical datasets for glioma patients are lacking; (ii) the white matter anatomy in glioma patients is often disrupted by tumor growth, which displaces surrounding tissues and deviates the bundle pathways from the canonical anatomy. This thesis presents two main contributions, each addressing one of these challenges. The first contribution addressed the issue of limited glioma clinical data by investigating the effectiveness of transfer learning. Specifically, we adapted a deep learning model, initially trained on a large healthy population, to a smaller dataset of glioma patients. We categorized the domain shift between healthy controls and glioma patients into two types: systematic domain shift (stemming from differences in acquisition and preprocessing) and tumor-related shift (arising from pathological anatomical changes). We developed metrics to quantify these shifts, allowing us to evaluate how transfer learning effectively solved systematic domain shift, but was insufficient in addressing tumor-related deviations. These findings were consistent across three imaging modalities and five major white matter bundles. The second contribution aimed to deal with the anatomical deviations in glioma patients by introducing a novel feature space, the Deformation Features, which encode patient-specific spatial relationships between the tumor and tractography. We used these features as input to a supervised geometric deep learning model designed to learn the geometric properties of fibers. Our results demonstrate that the model can successfully segment bundles even in the presence of tumor-induced anatomical alterations. Together, these two contributions provide significant advancements toward the automated segmentation of white matter tracts in the clinically relevant and complex context of glioma patients.