Click here to


Are you sure ?

Yes, do it No, cancel

Predicting Motion Compensation Errors for CyberKnife SBRT Liver Patients Using a Machine Learning Algorithm

M Liu1*, D Granville2 , J E Cygler1,2, E Vandervoort1,2 (1) Carleton University, Ottawa, ON (2) The Ottawa Hospital Cancer Centre, Ottawa, ON


(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Room 209

Purpose: To predict respiratory motion compensation (Synchrony) errors for CyberKnife treatments based on the analysis of treatment log files for the first fraction by training a support vector machine (SVM).

Methods: 14 features of the first fraction of treatment are extracted from CyberKnife treatment log files for 75 liver patients. External breathing motion is analyzed by decomposing the external marker motion into two principal components (PC1 and PC2). The peak frequency and signal-to-noise ratio (SNR) are extracted after the fast Fourier transform applied to PC1. Baseline shifts are calculated using PC2. Both linear discriminant analysis (LDA) and principal component analysis (PCA) are used to project the original features into a 2D space with two and three class labels, respectively. The class labels are based on the magnitude of Synchrony errors for the remaining fractions. The SVM classifier is trained using these transformed data and its quality is evaluated by means of a confusion matrix and 3-fold cross-validation.

Results: The 2D projections of all features using LDA with three class labels display three distinguishable clusters. A classification accuracy of 81% ± 5%, using 3-fold cross-validation, is achieved with the SVM. The 2D projections using PCA with two labels show two separable clusters (SVM classifier accuracy of 80% ± 6% using 3-fold cross-validation). The features derived from first fraction log files which are of importance in classifying the groups are predictor error, correlation errors, peak breathing frequency, linearity between internal and external breathing motion, baseline shift of external motion, and SNR in external breathing patterns.

Conclusion: The support vector machine is a feasible technique for classifying the SBRT liver patients with different Synchrony errors. Such a prediction could allow for adaptive adjustment of the PTV margins after the first fraction to improve tumor control and reduce treatment toxicity.

Funding Support, Disclosures, and Conflict of Interest: Funding for this research is partially provided by a grant from Accuray Inc. and The Ontario Consortium for Adaptive Interventions in Radiation Oncology.


Statistical Analysis, Stereotactic Radiosurgery, Respiration


TH- Dataset analysis/biomathematics: Machine learning techniques

Contact Email