Click here to

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

Automated Patient Positioning Error Detection with Orthogonal Setup DRR and Treatment KV Radiograph Image Pairs

J Charters*, R Petragallo, D Luximon, J Neylon, D Low, J Lamb, University of California, Los Angeles, Los Angeles, CA

Presentations

TU-D1030-IePD-F1-1 (Tuesday, 7/12/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 1

Purpose: In image-guided radiotherapy (IGRT), patient radiographic setup errors are rare but harmful if not corrected. Along the vertebral column it can often be difficult to discern from radiographs whether a patient is misaligned by one vertebral body. In this study, we investigate the use of convolutional neural networks (CNN) to assist in detecting vertebral body misalignments on orthogonal kV radiographs for application to retrospective error hunting and physicists’ weekly chart-check reviews.

Methods: 2672 images acquired from N=69 radiotherapy patients aligned with planar kilovoltage (kV) radiographs were used to develop an error-detection module. Digitally reconstructed radiographs (DRRs) and orthogonal kV images were retrieved and co-registered according to the clinically applied alignment contained in the DICOM REG files. A semi-automated algorithm was developed in MATLAB to simulate patient positioning errors on the anterior-posterior (AP) images shifted by one vertebral body along the superior-inferior (SI) direction. Registrations were achieved by maximizing zero-normalized cross-correlations (ZNCC) between the orthogonal DRR and radiograph pairs within a limited search window about an initial alignment point manually selected near an off-by-one vertebral body. Linear discriminant analysis (LDA) was performed to identify a linear classifier on the image gradients. Finally, a CNN architecture containing two convolutional and two dense blocks was designed to classify either AP images individually or AP and LAT images together. Receiver-operator characteristic curves (ROC) and areas under the curves (AUC) were computed to evaluate the classifiers.

Results: The AUC for our linear discriminant was 0.92. The AUCs for the CNN models were 0.97 for AP only and 0.95 for combined AP and LAT.

Conclusion: LDA provides a simple and effective method to classify patient setup errors, although our trained CNN outperformed linear discriminants. Our CNN error detection network successfully classifies vertebral body positioning errors with sufficient accuracy for retrospective quality control or chart-check reviews.

Funding Support, Disclosures, and Conflict of Interest: Supported in part by the Agency for Healthcare Research and Quality (AHRQ) under award number 1R01HS026486

Keywords

Setup Errors, ROC Analysis, Linear Discriminant Analysis

Taxonomy

TH- Dataset Analysis/Biomathematics: Machine learning techniques

Contact Email

Share: