Can Transfer Learning Improve Supervised Segmentation of White Matter Bundles in Glioma Patients?

In clinical neuroscience, the segmentation of the main white matter bundles is propaedeutic for many tasks such as pre-operative neurosurgical planning and monitoring of neuro-related diseases. Automating bundle segmentation with data-driven approaches and deep learning models has shown promising accuracy in the context of healthy individuals. The lack of large clinical datasets is preventing the translation of these results to patients. Inference on patients’ data with models trained on the healthy population is not effective because of domain shift. This study aims to carry out an empirical analysis to investigate how transfer learning might be beneficial in overcoming these limitations. For our analysis, we consider a public dataset with hundreds of individuals and a clinical dataset with glioma patients. We focus our preliminary investigation on the corticospinal tract and on the inferior longitudinal fasciculus. The results show that transfer learning is effective in overcoming part of the domain shift.