Program Information
Prediction of Laser Induced Thermal Therapy: Results of Model Training and Cross Validation
SJ Fahrenholtz1,2*, R Madankan1 , JD Hazle1,2 , RJ Stafford1,2 D Fuentes1,2 , (1) UT MD Anderson Cancer Center, Imaging Physics, Houston, TX, (2) UT Graduate School of Biomedical Sciences at Houston, Medical Physics, Houston, TX
Presentations
SU-C-BRA-3 (Sunday, July 12, 2015) 1:00 PM - 1:55 PM Room: Ballroom A
Purpose: MR-guided laser induced thermal therapy (MRgLITT) is a minimally invasive surgery with applications in the brain, among other sites. In especially precise interventions, like neurosurgery, accurate planning may behoove surgical planning by aiding in the decision of where and how many laser ablations are required. Previous models of tissue heating have relied on literature values extrapolated primarily from normal brain animal research and ex vivo data. In this abstract, an inverse problem provides model parameter data from retrospective analysis of MR temperature imaging data in patient tumor tissue, which represent a training cohort. Within the same cohort, leave-one-out cross validation (LOOCV) estimates the predictive accuracy of the trained model.
Methods: The training has three parts: MR temperature datasets (n=20), a relatively simple steady state bioheat model, and a global optimization algorithm maximizing the Dice similarity coefficient (DSC). DSC ranges from 0 to one; >0.7 is considered a ‘successful’ prediction. DSC compares the regions exceeding 57 °C—i.e. ablated tissue regions—and measures accuracy. Blood perfusion and optical parameters are optimized according to DSC, creating a library of 20 pairs of parameters that are used in prediction. The predictive accuracy is estimated by using LOOCV. LOOCV begins by dropping an optimal parameter pair from the library and makes a model prediction, given the average from the remaining 19 pairs. This procedure is permuted so that every dataset is predicted using the other parameter pairs.
Results: The distribution of DSC during predictive LOOCV is described by: mean=0.822; median=0.849; standard deviation=0.0872; minimum=0.583; maximum=0.930; 19/20 datasets pass (i.e. DSC>0.7). Best available literature values perform comparably worse: mean=0.668; median=0.656; standard deviation=0.115; minimum=0.455; maximum=0.857; 8/20 pass.
Conclusion: Data strongly indicate population-based parameter optimization is potentially useful in treatment planning, as the model far outperforms available literature values in this preliminary study.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this publication was supported by the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number TL1TR000369. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Contact Email: