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

Accurate prediction of the laser energy absorption and corresponding thermal spread is essential for safe and effective outcomes in magnetic resonance-guided laser interstitial thermal therapy (MRgLITT), as it enables clinicians to anticipate thermal spread and ensure complete ablation of the epileptogenic focus while minimizing collateral damage. However, current planning tools rely on simplified models that neglect patient-specific anatomy and the cooling effects of cerebrospinal fluid (CSF), often resulting in incomplete or asymmetric ablations. To address these limitations, this work sought to improve preoperative planning for MRgLITT by integrating the bioheat transfer equation (BHTE) with U-Net-based deep learning to predict patient-specific temperature maps.

We leveraged a large dataset of paired presurgical MR and intraoperative temperature maps from 340 MRgLITT procedures for mesial temporal lobe epilepsy. We proposed a hybrid model, physics-assisted U-Net (PA-U-Net) that used planning MRIs, tissue segmentations, and a physics prior to predict intraoperative temperature maps. We compared its performance with two baseline models: a physics-only model (BHTE) and a deep learning model (U-Net). All three models were evaluated against ground-truth magnetic resonance thermometry (MRT). We evaluated the spatial and morphological accuracy of an ablation zone, delineated by a 42 °C temperature threshold, using the Dice similarity coefficient and a roundedness (4πA/P²) metric, as well as pixel-wise root-mean-square error (RMSE, °C) and the percentage of well-predicted voxels (|prediction - ground truth|≤ 5 °C). A generalized linear model analysis was performed to evaluate how local tissue composition, particularly CSF proximity, influenced ablation geometry and model performance.

The models were trained on 313 MRgLITT cases and evaluated on 27 test acquisitions. The PA-U-Net achieved the highest overall spatial agreement with the ground-truth MRT thermal distribution, with a Dice score of 0.74, significantly higher than that of the U-Net (p < 0.001). Although PA-U-Net and BHTE showed similar overall Dice scores, threshold-wise analysis revealed that PA-U-Net consistently maintained superiority at higher isotherm thresholds, indicating improved localization of hotter ablation cores. Per-case analysis showed that PA-U-Net outperformed BHTE in 18 of 27 test cases, demonstrating stronger and more consistent performance across subjects. Roundedness analysis showed that both U-Net and BHTE differed significantly from the ground truth (p < 0.001), whereas PA-U-Net showed no significant difference (p > 0.05), indicating the closest alignment with the true ablation geometry. Ground-truth roundedness showed no significant dependence on CSF proximity across the full dataset (p = 0.23), but within the subset of cases where PA-U-Net outperformed both BHTE and U-Net, the relationship exhibited a marginal negative trend (p = 0.06), suggesting that higher CSF content around the laser tip was associated with lower roundedness. Pixel-wise analyses showed that PA-U-Net achieved an RMSE of 2.68 ± 0.47 °C and a well-predicted voxel percentage of 86%.

Embedding a physics-based heat prior into the U-Net, PA-U-Net achieved both anatomical adaptability and physical consistency, accurately reproducing temperature magnitude and spatial shape characteristics of true thermal distributions. This hybrid framework outperformed data-driven and physics-only models, particularly in complex anatomical regions where traditional physics-based methods fail. These results demonstrate that physics-assisted deep learning can substantially enhance MRgLITT treatment planning and lay the groundwork for future AI-assisted, patient-specific surgical planning tools.