Assessing informative tract segmentation and nTMS for pre-operative planning

Area of Science:

• Neuroimaging
• Neurosurgery
• Artificial Intelligence

Background:

• Deep learning (DL) excels at white matter tract segmentation in healthy subjects.
• Tract annotation is challenging in clinical data, especially in tumor patients with distorted anatomy.
• Direct cortical stimulation is invasive; navigated transcranial magnetic stimulation (nTMS) offers a non-invasive alternative for motor mapping.

Purpose of the Study:

• To evaluate nTMS motor responses for assessing DL-based corticospinal tract (CST) binary masks and uncertainty in diffuse glioma patients.
• To validate the feasibility of using nTMS as a ground truth for tract segmentation.

Main Methods:

• Utilized nTMS motor responses to assess CST binary masks and uncertainty from a DL-based segmentation method (UncSeg).
• Compared UncSeg's performance against the state-of-the-art TractSeg using overlap coefficient (OC) with nTMS response masks.

Main Results:

• CST binary masks generated by DL showed high overlap coefficients (OC) with nTMS response masks.
• A significant negative correlation was observed between segmentation uncertainty and the distance of nTMS response masks to the CST binary mask boundary.

Conclusions:

• Estimated uncertainty from UncSeg effectively measures agreement between CST binary masks and nTMS response masks.
• This approach provides a reliable, non-invasive method for validating CST segmentation in neuro-oncology patients.