Proof of concept of a fully unsupervised anomaly detection framework in CBCT-guided radiotherapy

Anomalies in cone beam computed tomography (CBCT) radiotherapy image guidance, such as setup misalignments and soft-tissue variations, can be indicative of treatment deviations potentially impacting quality and safety. Repetitive review of routine alignment images by human observers is inefficient, prone to cognitive biases, and poorly suited for the detection of rare events.

We propose an unsupervised image-guidance anomaly recognition and detection (iGuARD) framework, based on a CBCT inpainting technique using a variational autoencoder (VAE), which would highlight anomalies for human review.

The iGuARD framework was developed to output an anomaly score which would be highest for images containing infrequently observed, abnormal features. Algorithm training and testing data consisted of clinically registered simulation computed tomography (simCTs) and setup CBCTs from 1130 radiotherapy patients at the UCLA Medical Center. Using as input the simCT and the corresponding CBCT, both with two octants zeroed, the VAE was trained to inpaint the CBCT scan. The VAE's inpainting accuracy degrades in the presence of unusual image features, allowing the detection of anomalies through image similarity measures between actual and inpainted CBCTs. The VAE was subsequently applied to an unseen test dataset from 243 patients, including seven known misalignment incidents, 223 simulated 2 cm translational alignment errors, and 10 simulated wrong patient registrations. To assess the reconstruction accuracy of the inpainted CBCT, seven metrics were calculated based on the mean-square error, mean absolute error, mutual information, and structural similarity index measure. Principal component analysis was used to reduce the similarity measures' dimensionality to two, and a k-means clustering algorithm identified two clusters, of which the centroid (C_norm) of the denser cluster was extracted. The distance between each sample test point and C_norm was used as the anomaly score. A Receiver Operating Characteristic (ROC) curve was built to assess the algorithm's performance. For comparison, the experiment was repeated using the metrics obtained between the simCT and setup CBCT, excluding the VAE (traditional method).

For a fixed sensitivity of 95%, the specificities of the iGuARD framework and the traditional method were found to be 74.1% and 59.9% respectively on the unseen test dataset. When applied to all 1110 patients' data (i.e., whole dataset excluding the simulated errors), the iGuARD framework identified all seven known incidents (100% sensitivity) with a specificity of 93.0%, while the traditional method had a specificity of 77.9% for similar sensitivity. Upon review of the cases obtaining an anomaly score in the 99^th percentile range, we observed that those treatments often showed irregularities such as substantial soft tissue variations (e.g., different bladder and bowel filling affecting the target location) and subpar CBCT image quality.

The novel iGuARD framework presented in this study offers a way to automatically identify the patient setup CBCT scans and treatment fractions which are deviating from normality and may require the attention of the physician and/or physicist. This tool may not only improve the efficiency of repetitive and time-consuming quality assurance checks but may positively impact patient safety and treatment outcomes.