Patient-specific instantaneous spatial temperature maps for MR-guided laser interstitial thermal therapy using a...

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate thermal spread prediction is crucial for MR-guided laser interstitial thermal therapy (MRgLITT) outcomes.
• Current MRgLITT planning tools lack patient-specific anatomy and cooling effect integration, leading to suboptimal ablations.
• Improving preoperative planning is essential for safe and effective MRgLITT procedures.

Purpose of the Study:

• To enhance MRgLITT preoperative planning by integrating the bioheat transfer equation (BHTE) with U-Net deep learning.
• To predict patient-specific temperature maps for improved thermal spread anticipation.
• To address limitations of simplified models in MRgLITT planning.

Main Methods:

• Developed a hybrid physics-assisted U-Net (PA-U-Net) model using planning MRIs, tissue segmentations, and a physics prior.
• Compared PA-U-Net against physics-only (BHTE) and deep learning (U-Net) models.
• Evaluated model performance using ground-truth magnetic resonance thermometry (MRT) and metrics like Dice similarity coefficient and roundedness.

Main Results:

• PA-U-Net demonstrated superior spatial agreement with ground-truth MRT thermal distribution (Dice score 0.74), outperforming U-Net (p < 0.001).
• PA-U-Net showed no significant difference in ablation zone roundedness compared to ground truth, unlike U-Net and BHTE.
• Pixel-wise analysis showed PA-U-Net achieved an RMSE of 2.68 ± 0.47 °C and 86% well-predicted voxels.

Conclusions:

• The PA-U-Net model achieves anatomical adaptability and physical consistency, accurately reproducing thermal distributions.
• This hybrid framework outperforms traditional models, especially in complex anatomical regions.
• Physics-assisted deep learning offers a promising approach for AI-assisted, patient-specific surgical planning in MRgLITT.