Click here to

Session: Applications of AI in Radiotherapy Planning and Adaptation [Return to Session]

An Automated Pipeline for AI-Based Analysis of CBCT-Guided Patient Alignment

J Neylon1*, D Luximon1, T Ritter2, J Lamb1, (1) UCLA, Los Angeles, CA, (2) VCU Health System, Chesterfield, VA


SU-J-BRC-2 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: A pipeline was developed to supplement periodic review of pre-treatment image guidance, utilizing an in-house deep learning algorithm to flag potential misalignments for in-depth inspection.

Methods: Python scripts were developed to interact with the clinical database through DICOM networking protocol to automate data retrieval and analysis. Initial deployment of the automated pipeline was configured to execute nightly, and produce daily summary reports. Initial action levels targeted 10% of cases to be highlighted for further physics review. Highlighted cases were then reviewed by a physicist and classified.

Results: To date, the pipeline has analyzed a total of 487 CBCT registrations over 15 treatment days. Based on preliminary action levels, 34 (7%) cases from 23 patients have been flagged for closer inspection. Of these 34 cases, 21 could be generally classified as aligned to target/fiducial, with discernible differences in posture. 3 cases were attributed to changes in anatomy, such as weight loss, tumor shrinkage, or rectal filling. 5 cases were classified as ‘Conflicting Priorities’, where a compromise was deemed necessary between postural bony anatomy and the soft tissue targets. Lastly, 5 cases were flagged as False Positives. Of these 5 cases, 3 were attributed to preprocessing errors where the relevant anatomy was not properly captured for analysis. The two remaining false positives were judged properly aligned to the vertebral bodies on review. One case exhibited a difference of several centimeters in the superior edge of the liver, while the other had no obvious discrepancies.

Conclusion: Further data collection and validation is required, including an exhaustive search for false negatives and consideration of physician intent regarding image guidance. In addition, optimal action levels still need to be solidified, but early impressions show the potential benefit of automation and AI to increase the robustness of quality control in the radiotherapy treatment workflow.

Funding Support, Disclosures, and Conflict of Interest: The research reported in this study was supported by the Agency for Healthcare Research and Quality (AHRQ) under award number 1R01HS026486.


Cone-beam CT, Quality Control, Computer Vision


IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer/machine vision

Contact Email