Supervised White Matter Bundle Segmentation in Glioma Patients With Transfer Learning

Abstract

In clinical neuroscience, the segmentation of the main white matter bundles is a mandatory premise for pre-operative neurosurgical planning, neuro-rehabilitation prognosis and monitoring of neuro-related diseases. The automation of the task of bundle segmentation is achieving a good accuracy on healthy individuals taking advantage of data driven approaches and deep learning models. The lack of large clinical datasets is preventing the translation of these results to patients. Inference on patients data with model trained on healthy population is not effective because of domain shift. This study aims to carry out an empirical analysis to investigate how techniques of transfer learning might be beneficial to overcome these limitations. For our analysis, we consider an open publicly available dataset with hundreds of individuals and a clinical dataset with several tens of glioma patients. We focus our preliminary investigation on corticospinal tract. The results show that transfer learning might be effective to overcome the component of domain shift related to systemic bias. Nevertheless, the alterations induced by the tumors are not successfully managed by transfer learning and are still preventing the successful translation of model trained on healthy population.