Author: James M. Lamb, Dishane Chand Luximon, Jack Neylon, Rachel Petragallo, Moritz Ritter, Timothy Ritter 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles, ETH Zurich, VCU Health System, Department of Radiation Oncology, University of Colorado 🌍
Purpose: Anomalies in cone beam computed tomography (CBCT) radiotherapy image guidance can signal treatment deviations. Repetitive review of setup image registrations by humans is inefficient, prone to cognitive biases, and imperfect at identifying such rare events. We propose an unsupervised image-guidance anomaly recognition and detection (iGuARD) framework to highlight anomalies for human review.
Methods: iGuARD generates an anomaly score which is highest for images containing infrequently observed features. 9,488 clinically registered and physician-approved simulation computed tomography images (simCTs) and setup CBCTs were obtained from 1,055 patients treated at the UCLA Medical Center between 2016 and 2017. The dataset was randomly split 90:10 for training and testing based on patients' unique identifications. Using as input the simCT and the corresponding CBCT, both with two octants zeroed, a variational autoencoder (VAE) was trained to inpaint the missing octants of the CBCT. The VAE's inpainting accuracy degrades in the presence of unusual image features, allowing anomaly detection through image similarity measures between actual and inpainted CBCTs. iGuARD was subsequently applied to the test set containing 1,005 registrations, including seven known misalignment incidents and 114 simulated translational errors. A Receiving Operating Characteristic (ROC) curve was built to assess algorithm performance. For comparison, the experiment was repeated using the metrics obtained between the simCT and setup CBCT, excluding the VAE inpainting (traditional machine learning method).
Results: With a fixed sensitivity of 95% at detecting the misalignments, the specificities of iGuARD and the traditional method were found to be 91.7% and 82.1% respectively on the unseen test dataset. When applied to all 1,055 patients’ data (excluding the simulated errors), iGuARD identified all seven known incidents with a specificity of 93.6%, compared to 84.4% for the traditional method.
Conclusion: The novel iGuARD framework offers a way to automatically identify setup CBCT registrations that deviate from physician-approved standards.