SU-F-J-02: Flexible Training of MR-Guided Laser Ablation Models Via Global Optimization.

PURPOSE MR-guided laser ablation is currying interest in neurosurgery as a minimally invasive procedure. The procedure is planned via an assumed lesion overlaid on various anatomic and functional images. Computational models of heat transfer may further inform planning the intervention, reducing surgical morbidity and ensuring therapeutic effect. These models require physical parameter data, via less relevant literature values. Using commercialized laser ablation equipment combined with MR temperature imaging (MRTI) feedback, recent work has demonstrated the use of inverse problems as a means to gain the necessary information via retrospective analysis. Here, performance of a global inverse problem solution is explored that allows for simultaneous optimization of multiple objective functions. METHODS A steady state bioheat model is trained on MRTI patient datasets (n=20); the MRTI regions exceeding a 57°C threshold are considered lethally damaged. Optical and blood perfusion parameters are densely sampled, allowing for global optimization of the model. There are two objective functions to optimize with respect to damage prediction: Dice similarity coefficient (DSC) and Hausdorff distance (HD). DSC is an overlap metric; HD is the distance-to-agreement in physical distance. Predictive performance is simulated via leave-one-out cross-validation (LOOCV) and is compared to the same model using naive literature values. RESULTS Post-optimization and during LOOCV, mean and median DSC are 0.824 and 0.850 respectively; mean and median HD are 4.64mm and 4.03mm respectively. These compare favorably to the model using literature values. Mean and median DSC are 0.6685 and 0.6564 respectively; mean and median HD are 5.85mm and 5.58mm respectively. A two-tailed paired Student's t-test for the mean DSC reported p=0.0002. For mean HD, p=0.0064. CONCLUSION Clinical MRTI feedback provides useful information for designing optimized laser ablation models. Global optimization facilitates simultaneous optimization of multiple objective functions. Trained models are considerably more accurate than naive literature based models as measured by DSC and HD.