Encrypted login | home

Program Information

Knowledge-Based Quality Control of Clinical Stereotactic Radiosurgery Treatment Plans


S Shiraishi

S Shiraishi1*, J Tan2 , L Olsen2 , K L Moore1 , (1) University of California, San Diego, La Jolla, CA , (2) Washington University in St. Louis, St. Louis, MO

Presentations

TH-A-9A-8 Thursday 7:30AM - 9:30AM Room: 9A

Purpose: To develop a quality control tool to reduce stereotactic radiosurgery (SRS) planning variability using models that predict achievable plan quality metrics (QMs) based on individual patient anatomy.

Methods: Using a knowledge-based methodology that quantitatively correlates anatomical geometric features to resultant organ-at-risk (OAR) dosimetry, we developed models for predicting achievable OAR dose-volume histograms (DVHs) by training with a cohort of previously treated SRS patients. The DVH-based QMs used in this work are the gradient measure, GM=(3/4pi)^1/3*[V50%^1/3–V100%^1/3], and V10Gy of normal brain. As GM quantifies the total rate of dose fall-off around the planning target volume (PTV), all voxels inside the patient’s body contour were treated as OAR for DVH prediction. 35 previously treated SRS plans from our institution were collected; all were planned with non-coplanar volumetric-modulated arc therapy to prescription doses of 12-25 Gy. Of the 35-patient cohort, 15 were used for model training and 20 for model validation. Accuracies of the predictions were quantified by the mean and the standard deviation of the difference between clinical and predicted QMs, δQM=QM_clin–QM_pred.

Results: Best agreement between predicted and clinical QMs was obtained when models were built separately for V_PTV<2.5cc and V_PTV>2.5cc. Eight patients trained the V_PTV<2.5cc model and seven patients trained the V_PTV>2.5cc models, respectively. The mean and the standard deviation of δGM were 0.3±0.4mm for the training sets and -0.1±0.6mm for the validation sets, demonstrating highly accurate GM predictions. V10Gy predictions were also highly accurate, with δV10Gy=0.8±0.7cc for the training sets and δV10Gy=0.7±1.4cc for the validation sets.

Conclusion: The accuracy of the models in predicting two key SRS quality metrics highlights the potential of this technique for quality control for SRS treatments. Future investigations will seek to determine whether QM variations are due to residual model inaccuracies or true plan quality variations in the clinical sample.


Funding Support, Disclosures, and Conflict of Interest: Support from Varian Medical Systems.


Contact Email: