Click here to


Are you sure ?

Yes, do it No, cancel

Predicting Individual Target V12Gy for Single Isocenter Multi-Target (SIMT) Stereotactic Radiosurgery(SRS): A Comparison of Three Empirical Methods

T Li1*, H Zhang, W Shi2 , H Liu2 , (1) University of Pennsylvania, Philadelphia, PA, (2) Thomas Jefferson University, Philadelphia, PA


(Saturday, 3/30/2019)  

Room: Exhibit Hall

Purpose: V12Gy is the key indicator for brain necrosis following SRS. This study compares three empirical methods for predicting V12Gy from target volume and prescription information prior to planning.

Methods: 16 patients with a total of 112 targets was retrospectively studied. Plans were generated using Brainlab Elements™ automatic SIMT system. Three methods were compared for predicting V12Gy in cc using only the target volume (cc) and prescription (Gy) information. (1) Linear regression based on individual target volume weighted by prescription. (2) Neuro-network-based regression using 10 hidden layers and Bayesian Regularization backpropagation training algorithm. The training was performed in MATLAB® with 85% of samples randomized to training set and 15% as test set. To assess the randomness of training/testing partitioning, the training was repeated 100 times, and the Root-mean-square-error (RMSE) of test data prediction was summarized. (3) Geometric expansion (1-6mm) of the target’s equivalent sphere by a fixed distance to approximate the location of 12Gy isodose volume outside the target. The volume of this expanded target was then compared to true V12Gy to determine residual prediction error.

Results: (1) Multivariate analysis showed that V12Gy is highly correlated to the product of target volume (TV) and its prescription dose (Rx). Linear regression between V12Gy and achieved an RMSE of 1.09 cc. (2) Neuro-network-based regression was performed 100 times and showed RMSE of 1.14±0.04 cc (range 1.026-1.336). (3) Residual analysis showed that a fixed expansion of 4 mm achieved smallest and most stable prediction with RMSE at 1.12 cc.

Conclusion: Despite more complicated algorithm and 3D geometric information that were included in methods 2 and 3, simple linear regression achieved the lowest RMSE. It is feasible to predict individual target’ V12Gy during SIMT SRS using simple linear regression based on target volume weighted by prescription with an RMSE of ~1.1 cc.

Contact Email