(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To develop a deep-learning-based model to predict achievable DVHs of organs-at-risk (OARs) for automation of inverse planning.
Methods: The model was based on a connected residual deconvolution network. The contours of PTVs and OARs were parsed from plan database and taken as the input of the model, and the OARs’ dose area histograms (DAHs) derived from DVHs were taken as the output of the model. The model was trained from scratch by correlating anatomical features with OARs’ DAHs, which were accumulated to get the final predicted DVHs of each OAR. Helical tomotherapy (HT) plans for 190 nasopharyngeal cancer (NPC) patients were used to train, validate and test the model. The DVHs and specific dose-volume indices (DVIs) predicted from the model were compared with the clinical DVHs and DVIs from each patient’s plan in the testing dataset. The mean absolute errors (MAE) of DVI for each OAR were calculated and statistically analyzed with the Paired-samples T Test.
Results: Dose volume histograms of twenty-one OARs of nasopharyngeal cancer have been predicted by the CResDevNet model. For each OAR, the DVIs of interest compared were different combinations of V20, V30, V40, V50, D2% and mean dose (Dmean). For each testing patient, sixty-three DVIs for all OARs were calculated. For twenty testing patients, we found that 78% or 94% DVIs of all OARs were within 5% or 10% error. The MAE were ranged from 0.7±0.5% to 9.7±7.6% with median 3.1±2.3% for DVIs of all OARs.
Conclusion: This study developed a deep learning model for predicting achievable OARs’ dose volume histograms. The prediction accuracy was evaluated with nasopharyngeal cancer cases, and shown to be more accurate for large OARs than for small OARs. The model is useful in automation of inverse planning.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
