Prediction of Laser Ablation in Brain: Sensitivity, Calibration, and Validation

The surgical planning of MR-guided laser induced thermal therapy (MRgLITT) stands to benefit from predictive computational modeling. The dearth of physical model parameter data leads to modeling uncertainty. This work implements a well-accepted framework with three key steps for model-building: model-parameter sensitivity analysis, model calibration, and model validation. The sensitivity study is via generalized polynomial chaos (gPC) paired with a transient finite element (FEM) model. Uniform probability distribution functions (PDFs) capture the plausible range of values suggested by the literature for five model parameters. The five PDFs are input separately into the FEM model to gain a probabilistic sensitivity response of the model to the input PDFs. The result demonstrates the model output variance is dominated by the three optical parameters and the two remaining parameters contribute less. The second aim is model calibration, given the need to acquire model parameter data of greater precision sans physical measurement. The availability of a relatively large cohort of N = 22 clinical laser ablations of metastases gradient-based inverse problems provides inference of the optical parameter values, the most sensitive parameter as indicated by gPC, from patient MR temperature imaging (MRTI). In order to accelerate the bioheat model for iteration during parameter optimization, two simplified models are conceived: (1) a homogeneous, transient FEM model implemented on GPU and (2) a homogeneous, steady-state, analytic model implemented on GPU. After model optimization — i.e., calibration — the model validation immediately follows via leave-one-out cross-validation (LOOCV). LOOCV compares the two trained models’ predictive performances. During LOOCV, the FEM model correctly predicts 15 of 22; the steady state model correctly predicts 17 of 22. A steady state model using na¨ıve literature values correctly predicts only 10 of 22. When training on an N = 20 cohort tailored to only include ablations near steady state, the trained steady state model correctly predicts 19 of 20 patient datasets versus the 8 of 20 predicted by an untrained steady state model. The conclusion is model training is an effective means of improving model performance when there is lack of accurate and precise parameter data in the literature, especially when there is little prospect of improving data quality. A key to success in this model-training paradigm is to have a training/calibration cohort that has adequate similarity to the predicted/validation cohort