(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Objective: To assess the dosimetric robustness of delineations created via OARnet in comparison with UaNet for head and neck (H&N) organs-at-risk (OARs) in radiation therapy treatment plans.
Method/Materials: A new compact 3D deep learning model architecture (OARnet) and the publish UaNet method are used to delineate OARs in 28 publically available H&N CT datasets. For each method, delineation variabilities (DVs) were simulated by using the network to obtain OAR occupancy probability maps, then masking each probability map at seven different threshold levels to achieve different delineation samples. Dose distributions provided with the H&N CT sets were used to assess the dose-volume histograms (DVHs) for each sample and each network as well as with the reference OARs which were manually delineated by trained radiation oncologists. The accuracy of each model’s DVHs were assessed with reference to the manual delineation OARs. The DVH sensitivity to the delineation model was assessed by intercomparing the DVHs generated at each threshold for each model.
Results: DVHs from OARnet delineations were closer to the manual delineation DVHs than the UaNet delineation DVHs. OARnet delineations had a narrower 95 percent DVH confidence bands than UaNet delineations, indicating that the OARnet architecture is more robust to probability map threshold values
Conclusion: For radiation therapy, a delineation’s required accuracy is with reference to the dose distribution. DVH analysis reveals that OARnet DVHs more authentically reproduce manual delineation DVHs than the state-of-the-art UaNet-based delineations.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01CA222216.
Keywords
Dose Volume Histograms, Dosimetry, CT
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email