Machine learning model for predicting interfraction motion of the seminal vesicles in prostate cancer radiotherapy.

[BACKGROUND AND PURPOSE] In external beam radiotherapy for prostate cancer, inclusion of the seminal vesicles (SV) in the clinical target volume (CTV) is often complicated by considerable SV motion and deformation. This study aimed to investigate the feasibility of predicting patient-specific SV motion using anatomical features surrounding the prostate on planning CT (pCT) images.

[MATERIALS AND METHODS] Interfractional SV motion was quantified using five pretreatment cone-beam CT (CBCT) scans per patient from a cohort of 191 prostate cancer patients. Patients whose SV was not fully covered by a 3-mm margin were assigned to the High SV Motion Group, which served as the target for prediction. A total of 42 anatomical features were extracted from the contours of the prostate, SV, bladder, and rectum on the pCT. Feature selection was performed using Random-Forest Recursive Feature Elimination, and a machine learning model was developed and evaluated using both internal and external patient cohorts.

[RESULTS] Four anatomical features were selected, including those based on the anatomical relationship between the prostate and the SV. Using these features, the best-performing light gradient boosting machine model achieved an area under the receiver operating characteristic curve of 0.724 in the internal test and 0.632 in the external test for identifying patients in the High SV Motion Group.

[CONCLUSION] This study suggests an association between anatomical features derived from pCT and patient-specific SV motion. Although the current predictive performance is moderate, this approach may help support radiotherapy strategies when the SV is included in the CTV.