Click here to

Session: Patient Safety and Quality Improvement [Return to Session]

Automatic Detection of Vertebral Body Misalignment Errors in Cone-Beam CT Guided Radiotherapy Using a Convolutional Neural Network

D Luximon*, J Neylon, J Lamb, University of California, Los Angeles, Los Angeles, CA


SU-D-TRACK 6-3 (Sunday, 7/25/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Purpose: Vertebral misalignments during cone-beam computed tomography (CBCT)-guided radiotherapy can lead to wrong-site treatment. We propose an automatic error detection algorithm that uses a three-branch convolutional neural network (CNN) to detect vertebral body misalignments using planning computed tomography (CT) images and setup CBCT images.

Methods: Planning CTs and setup CBCTs were collected from 160 patients undergoing radiotherapy treatment in the thoracic or abdominal region using Varian TrueBeam and NovalisTx treatment machines. To simulate vertebral body misalignment errors, the setup and planning images were misaligned by one vertebral body and for each patient, one superior and one inferior misalignment were generated. The clinically applied registration was used to derive true-negative (no error) data. For each of the aligned and misaligned 3D image pairs, 2D slice pairs were extracted in each anatomical plane about an isocenter within the vertebral column. The 3 slice pairs obtained were then used as inputs to a three-branch CNN which returned a probability of vertebral misalignment. Each of the 3 branches was used to extract features from one of the orthogonal image planes, which were then joined in a single densely-connected layer. The CNN was trained on 140 patients and tested on 20 patients. The algorithm performance was quantified using a receiver operating characteristic (ROC) analysis. Due to the rarity of vertebral body misalignments in the clinic, a minimum threshold value yielding a specificity of at least 95% was selected.

Results: The ROC analysis resulted in an area under curve of 0.985 with a 95% confidence interval of [0.978 0.992]. The sensitivity was found to be 90% for a fixed specificity value of 95%.

Conclusion: Our CNN-based algorithm can accurately identify off-by-one vertebral body misalignments in planning CT and setup CBCT registrations, leading towards the possibility of an on-line error prevention system.

Funding Support, Disclosures, and Conflict of Interest: Research reported in this abstract was supported by the Agency for Healthcare Research and Quality under award number 1R01HS026486.



    Setup Errors, Image Guidance, Radiation Therapy


    IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

    Contact Email