Informative and Reliable Tract Segmentation for Preoperative Planning.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Manual white matter (WM) tract annotation for surgical planning is time-consuming and variable.
• Diffusion MRI noise and inter-rater variability complicate accurate WM tract identification.
• Direct electrical stimulation is often required during surgery to precisely locate WM tracts.

Purpose of the Study:

• To develop a deep learning model for reliable WM tract segmentation.
• To quantify segmentation unreliability using uncertainty estimation.
• To improve preoperative surgical planning and intraoperative guidance for WM tract localization.

Main Methods:

• Utilized a 3D U-Net architecture for WM tract segmentation.
• Employed test-time dropout and test-time augmentation for model and data uncertainty estimation.
• Applied a volume-based calibration approach to compute representative predicted probabilities.

Main Results:

• Achieved a Dice score of ≈0.82, comparable to state-of-the-art multi-label segmentation.
• Obtained a Hausdorff distance <10mm, indicating high spatial accuracy.
• Demonstrated a strong correlation between volume variance and segmentation errors, validating reliability estimation.
• Showed that calibrated predicted volumes better encompass ground truth segmentations.

Conclusions:

• Deep learning with uncertainty quantification offers a reliable method for WM tract segmentation.
• The developed approach provides a crucial measure of segmentation reliability for clinical use.
• This technique represents a significant advancement toward more informed and dependable WM tract segmentation in neurosurgery.