Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Improving Three-Dimensional Dose Prediction with Ordinal Regression

Y Yuan*, T Tseng, Y Lo, The Mount Sinai Hospital, New York, NY


PO-GePV-M-2 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To investigate the feasibility of leveraging the ranking information among the dose values in different voxels in the task of improving the three-dimensional dose prediction using deep neural networks.

Methods: We used 140 prostate cancer patients in this study, where 100 cases were used for model training and 20 for validation, and the rest 20 were reserved for independent model testing. All the patients were planned with RapidArc to prescription dose of 45 Gy. The volumetric dose prediction model was based on ResSE-UNet, where we replaced the convolutional blocks with residual network blocks and added squeeze-and-excitation (SE) module to further improve the representative capability of the model. The input to ResSE-UNet included the simulation CT and contours of several Organs-At-Risk (OARs), and the output was the predicted three-dimensional (3D) dose distribution. Considering that the dose values have a strong ordinal correlation since they form a well-ordered set, we explicitly modeled this ordering information by introducing an ordinal regression loss in addition to the mean absolute error (MAE) in the loss function for network training. The prediction accuracy was evaluated by comparing the clinical plan with the predicted dose distribution.

Results: On the testing dataset, the average predicted dose on PTV was found to be 46.72±0.92 Gy while the clinical dose was 46.81±0.78 Gy. The average voxel-wise dose absolute difference between the predicted and clinical dose distributions was reduced from 1.72 Gy to 1.61 Gy by incorporating the proposed ordinal regression loss, yielding 6.4% improvement.

Conclusion: Our preliminary results demonstrate that incorporating ordinal regression into the loss function can potentially improve the dose prediction accuracy, owing to its ability of explicitly modeling the ordered information between dose values. We are further optimizing this framework and evaluating its generalization capacity in more heterogeneous environment.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by a research grant from Varian Medical Systems (Palo Alto, CA, USA), UL1TR001433 from the NCATS/NIH and R21EB030209 from the NIBIB/NIH. The content is solely the responsibility of the author and does not necessarily represent the official views of the NIH.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email