Click here to

Session: Imaging for Patient Setup and Target Alignment [Return to Session]

Facilitation of Automated Estimation of Setup Uncertainty in Pediatric Proton Therapy Patients with Deep Learning and Optical Character Recognition

J Becksfort*, A Saunders, M Marker, T Russell, T Merchant, C Hua, St. Jude Children's Research Hospital, Memphis, TN


TU-D1030-IePD-F1-2 (Tuesday, 7/12/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 1

Purpose: To describe an automated data processing pipeline and report findings from a prospective trial quantifying setup uncertainty in children receiving proton therapy, including manual vs. cone-beam CT (CBCT) guided positioning, intra-fractional patient movement, and planar vs. volumetric image guidance.

Methods: Protocol-specified imaging included daily pretreatment CBCT, weekly post-treatment CBCT, verification CBCT after implementing setup corrections with a 6-degree-of-freedom robotic couch, and planar radiographs for comparison. Data for 169 (102 anesthetized) patients aged 3 months to 21 years were analyzed. Screenshots of image registration software displaying calculated setup corrections after manual positioning (n=4,885), after corrections implemented (n=237), and after treatment (n=677) were captured. A pipeline detected the screenshots, separated them into appropriate cohorts using TensorFlow, and used Tesseract to extract patient position coordinates. Results were stored in a PostgreSQL database and a web application facilitated the verification or correction of extracted data.

Results: Setup uncertainty for the cranial cohort after manual positioning was estimated (95th percentile) to be below 0.43 cm, 0.51 cm, 0.96 cm, 2.30°, 2.30°, and 2.20° for lateral, longitudinal, vertical, pitch, roll, and yaw directions, respectively. After image guided correction, the uncertainties were 0.090, 0.109, 0.060, 0.600°, 0.595°, and 0.495°. After treatment, uncertainties were 0.13, 0.14, 0.10, 1.00°, 0.80°, and 1.07°. Uncertainties for the body cohort were slightly larger. Pipeline accuracy for image selection and cohort classification was 98.31%.

Conclusion: Daily CBCT image guidance reduced setup uncertainty to within 1.1mm/0.6° and 1.43 mm/0.6° for the cranial and body cohorts, respectively. For anesthetized children, uncertainty remained small until the completion of treatment fractions. Study findings will guide planning margin design, robust optimization settings, and immobilization improvement. Automated system integration, cohort classification, and character recognition combined with a review interface improved the accuracy and efficiency of this data intensive research protocol.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email