Author: Yasin Abdulkadir, John Charters, Melissa Ghafarian, James M. Lamb, Dishane Chand Luximon, Jack Neylon 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles 🌍
Purpose:
Assessment of radiotherapy treatment quality in large-scale multi-institutional contexts remains an outstanding challenge. Retrospective human review of treatment plans is labor intensive and nearly impossible for large datasets. Dosimetric plan quality assessment can be approached with knowledge-based planning. However, a standardized, automated approach is lacking for other aspects of treatment quality. In this work, we develop novel automated methods for assessing plan quality indicators over large databases of prostate radiotherapy treatments. The purpose of the algorithms is to flag potentially low-quality plans for detailed human review.
Methods:
A DICOM database of prostate radiotherapy patients was analyzed, comprised of 2394 internal plans from our institution and 96 external plans. A convolutional neural network (CNN) was used to classify plans according to intact prostates or post-prostatectomy beds. Clinical target volumes (CTV) and planning target volumes (PTV) were automatically identified according to structure naming conventions and structure types. PTV margins were automatically derived by first maximizing a similarity coefficient with morphological expansions of the CTV, then computing a mean surface distance. The presence of fiducial markers was determined by intensity thresholding and voxel clustering. Treatment dose to the prostate was obtained by computing dose-volume histograms and interpolating the D95 point. Target contour outliers were identified according to geometric shape descriptors. To evaluate the accuracy of our models on the external plans, ground truth quality measurements were recorded manually.
Results:
Plan type classification accuracy was 93%. PTV and CTV identification accuracy was 96% and 97%, respectively. Significant aggregate differences in fiducial marker usage, PTV margins, and target dose were observed between internal and external plans. Geometric outliers correctly identified subjectively unusual contours.
Conclusion:
Automated algorithms for quantitative measurements of prostate plan quality were developed, and high accuracy was achieved on external data. Our methodology is pertinent to analyzing large radiotherapy databases.