Purpose: The radiation oncology clinical workflow involves numerous patient safety checks; however, rare but serious treatment errors do still occur and may go unreported. Here we present a semi-automated process, based on a convolutional neural network (CNN) misalignment detection algorithm, to retrospectively search a setup imaging database for near miss events and treatment errors not previously reported.
Methods: A CNN-based error detection model was developed using 57,036 digitally reconstructed radiograph (DRR)/ x-ray image pairs obtained from the records of patients treated at our institution from 2014-2017 using a stereoscopic onboard imaging guidance system. The model was trained to detect alignment errors between the x-ray and the DRR using simulated 1 cm translational errors. The model was applied to previously unseen data from 2018 in order to flag potential errors. Flagged events were cross-referenced with record-and-verify systems, incident learning systems, and external image-guidance systems to determine whether they represented treatment errors, near misses, or false positives.
Results: When the CNN was used to classify DRR/x-ray image pairs from 2018, a total of 1,233 image pairs (out of 27,278 total) were identified as misaligned. 136 image pairs were manually reviewed. Cross-referencing led to the following categorizations: treatment error, near miss, alignment mismatch corrected with repeat imaging, alignment mismatch corrected by imaging with a different system, fiducial marker match with bony anatomy mismatch, image acquisition failure, and algorithmic false positives.
Conclusion: Manual retrospective audits of image guided radiotherapy databases could lead to detection of previously unreported treatment errors, but are infeasible due to the sheer size of such databases. Previously unreported errors and near misses can provide key insights to improve the safety of the radiotherapy workflow. To our knowledge this report is the first use of an automated misalignment detection algorithm to pre-select potential errors, thus making systematic comprehensive retrospective review feasible.
Funding Support, Disclosures, and Conflict of Interest: This work supported in part by AHRQ R01 HS026486.
Not Applicable / None Entered.