Towards Knowledge Based Planning for CyberKnife® Stereotactic Radiosurgery Treatments of Multiple Intracranial Metastases

Abstract

Background: Stereotactic Radiosurgery (SRS) is used to treat both solitary and multiple Brain Metastases (BM). Improvements in treatment efficacy for primary tumours increases BM incidence, placing greater demands on centres delivering SRS. Knowledge Based Planning (KBP) has been demonstrated as a method to reduce the time required for radiotherapy planning, and so may help elevate the additional burden on SRS services. Methods: Planning data from a dedicated SRS platform (Accuray CyberKnife®) audit database was used to train and assess a series of KBP models, predicting the volume of whole brain (excluding Planning Target Volume; PTV) receiving radionecrosis tolerance doses. Models were taken from literature, including Bohoudi et al. (2016), Cummins et al. (2020) and Yu et al. (2021). Additional models based on a power law dose fall-off from the lesion centre were also explored, including a power law shifted away from the PTV centre. 289 single and 381 multi-lesion cases were used for training, with 122 and 259 independent cases for testing. Assessment included 10-fold cross validation on the training data, for a range of error metrics such as the Mean Absolute Percentage Error (MAPE). Model parameters learnt on single lesion cases were also applied to multiple BM cases; and finally voxelised variants of the power law models were created to assess the interaction between lesions. Results: For single lesions the MAPE from the training data cross validation (mean ± standard deviation) was 29 ± 6.9%, 23 ± 5.3%, 23 ± 6.0% and 22 ± 5.9% for Bohoudi et al. (2016), Cummins et al. (2020), Yu et al. (2021) and the shifted power law respectively; with errors of 37%, 22%, 21% and 20% for the test data. For multiple lesions, the MAPE was slightly lower at 20 ± 3.5%, 20 ± 3.8% and 20 ± 3.7% for Bohoudi et al. (2016), Yu et al. (2021) and the shifted power law; and 24%, 23% and 23% for the test data. The voxelised shifted power law model for multiple lesions had a MAPE of 18 ± 2.4% and 19% on the training and test data respectively. For some methods large errors occurred when applying single lesion parameters to multiple BM cases; for Bohoudi et al. (2016) the median MAPE [95% Confidence Interval] was 38 [32, 45]% for single lesion model coefficients compared to 23 [19, 26]% for multiple lesion learnt parameters. Conclusions: Separate models are required for both single and multiple lesion cases for CyberKnife treatments. The improvement derived from including lesion separation for these simple models is not sufficient to warrant the extra time required to acquire the information. Power law dose fall-off models were equivocal with other methods for CyberKnife, adding to existing evidence for this type of relationship, but further investigation is required.

Date of Award	31 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Julia-Claire Handley (Supervisor)