Using deep learning generated CBCT contours for on-line dose assessment of prostate SABR treatments




Smith, Conor

Journal Title

Journal ISSN

Volume Title



Prostate Stereotactic Ablative Body Radiotherapy (SABR) is a hypofractionated treatment regime where small errors in patient setup have the potential to deliver much higher doses to Organs at Risk (OARs). Inter-fraction variability is reduced by having patients follow a bowel and bladder preparation protocol, however patient adherence is inconsistent and OAR variability is still present. At our centre, radiation therapists use non-dosimetric decision trees (DT) to determine if the patient is sufficiently prepared and positioned, however the application of these decision trees varies between therapists. We used deep neural-network-generated cone-beam computed tomography (CBCT) contours to estimate the daily dose delivered to the rectum and bladder. We compared these estimates to the planned dose-volume metrics to facilitate future development of personalized and quantitative decision trees for prostate SABR treatments. Two-hundred pre-treatment CBCT scans from 40 prostate SABR patients, each prescribed 40 Gy in five fractions, were contoured using Limbus Contour AI based segmentation software (v1.5.0-D2). Dose-volume histogram (DVH) curves for each fraction’s OARs were calculated by combining the planned dose distribution and AI-generated contours using on-line rigid registration data. To evaluate the performance of patient bowel & bladder preparation and SABR decision trees, we calculated the same dose-volume metrics that were used as goals for treatment planning using the DVH curves generated from the daily CBCT contours. CBCT dose-volume metrics that met the preferred goal were considered as having ‘no violations’, and metrics that did not meet the mandatory goals were classified as ‘major violations’, with values in between categorized as ‘minor violations.’ Twenty-seven percent of fractions had at least one major violation in one of the bladder planning goals, and 34% had minor violations. Fourteen percent of fractions had a rectum metric major violation, with 10% categorized as minor violations. Across their five fraction treatment, five patients had recurring bladder V37 Gy major violations, and 2 patients had rectum V36 Gy major violations. Bowel and bladder preparation played a significant role as OAR differences in position and volume led to many instances of mandatory goals not being met. These results show a clear need for the development of rigorous and personalized decision trees. Online quantitative OAR volume data incorporated into a future DT would likely make it successful.



medical physics, radiation therapy, prostate SABR, autocontouring, organ at risk, organ at risk variability