Supervised white matter bundle segmentation in glioma patients with transfer learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Virtual dissection of white matter tracts aids neurological monitoring and treatment planning.
• Deep learning models show high accuracy for white matter segmentation in healthy individuals.
• Limited clinical datasets and population differences hinder deep learning application in patients.

Purpose of the Study:

• Investigate transfer learning for adapting deep learning models from healthy populations to glioma patients.
• Characterize domain shift components (systematic vs. tumor-specific) in white matter segmentation.
• Evaluate transfer learning's effectiveness in addressing data scarcity and domain shift.

Main Methods:

• Trained a deep learning architecture on a large healthy population dataset.
• Applied transfer learning to adapt the model for glioma patient data.
• Characterized domain shift by distinguishing systematic and tumor-specific components.
• Evaluated model performance across five white matter bundles and three input modalities.

Main Results:

• Models trained on healthy data showed a significant performance drop in glioma patients.
• Transfer learning effectively managed systematic domain shift and mitigated data scarcity.
• Fine-tuning could not fully compensate for large white matter deformations caused by tumors.
• Results demonstrated robustness and generalizability across different bundles and modalities.

Conclusions:

• Transfer learning is a viable strategy for white matter segmentation in clinical populations, especially for managing systematic domain shift.
• Deep learning models require further development to handle complex tumor-induced anatomical changes.
• This study provides insights for advancing automated segmentation and clinical transfer learning applications.