Program Information
Detection of OAR Segmentation Errors Using a Novel Automated QA Program
C Hui*, H Nourzadeh , B Neal , W Watkins , J Siebers , University of Virginia Health System, Charlottesville, VA
Presentations
SU-F-FS4-4 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: Four Seasons 4
Purpose: To uncover segmentation errors in a population of historic segmentation data set using an organ at risk (OAR) segmentation quality assurance (QA) algorithm.
Methods: The QA algorithm utilizes statistical anomaly detection technique to identify abnormal segmentation features. For each segmentation, 25 volumetric features are computed for comparison. The algorithm identifies features that are outside of the 95% interval with respect to historic distributions. Based on combinations of abnormal features, a multi-criteria triggering system warns the user of specific segmentation abnormalities. The program was tested on segmentations from 128 image sets of head and neck cancer patients. Feature distributions were first derived from the test sets. The algorithm was run on the same test sets and provided warnings to all abnormal segmentations. Guided by the warnings, reviewers identified and corrected for the segmentation errors. The frequency of error and its impact on the OAR structures were subsequently evaluated.
Results: The test sets contained 1156 OAR structures. Overall, we found 64 errors related to the extent of segmentation boundary, 12 structures with extra disconnected segments, 8 structures with missing slices, 7 wrong structures noticeably far away from the actual structures, and one case of switching left and right eye. Overall, the occurrence frequency of segmentation error was 8%. On average, boundary error resulted in 38.5% change in structure volume, whereas wrong segmentation led to 31.8% change in structure volume. In addition to volume uncertainty, wrong segmentation also caused an average shift of 1cm on the corresponding OAR structure. Mislabeling on the two eyes did not change their corresponding volumes but resulted in a 6cm shift in their positions.
Conclusion: The QA algorithm could help detect segmentation error in large scale data sets. In addition, it can be used to check new patient OAR structures with respect to the prior clinical standards.
Contact Email: