Click here to


Are you sure ?

Yes, do it No, cancel

Prediction of 3D Dose Distribution and PTV Contour Via Deep Multi-Task Learning Network

F Guo1*, F Kong1 , Y Li2 , L Zhou1 , T Song1 , (1) Southern Medical University, Guangzhou, Guangdong, (2) SUN YAT-SEN UNIVERSITY CANCER CENTER, Guangzhou, Guangdong,


(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: To build a multitask deep convolution network to predict both the 3D Dose distribution and outline PTV contour simultaneously for IMRT planning.

Methods: We collect the data from the clinical radiotherapy plan of the same tumor type , and learn our targets from extracted contour of OAR and the CT image. The deep work is made up of two parts. The first parts is constructed by a classic network architecture called U-net,which is designed for biomedical image segmentation. Here U-net is considered as a shared layer for common feature extraction. And the second part is two independent CNN network for dose distribution and PTV contour regression. This two regression module learning alternately to improve the performance without having to find more task-specific training data.To test the feasibility of this model, a total of 32 cases of Prostate cancer plans were randomly selected from clinical planning, 29 of which were used as model training, and 3 were model tests.

Results: 3D dose distribution prediction accuracy was presented with a mean dose error and standard deviation of 3 evaluated case is 4.58Gy±1.26. And the mean and standard deviation of contact ratio of PTV contour’s coincidence situation,which is calculated by the percentage of the PTV volume with coincidence voxel, is 0.87±0.09.

Conclusion: We successfully build a multitask deep convolution network model with the ability to predict voxel-level dose and meanwhile outline PTV contour.Besides it can also analyze image feature automatically by avoiding the artificial select geometric anatomical structure feature.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by (1) National Key R&D Program of China (NO.2017YFC0113200), (2) National Natural Science Foundation of China (NO.81571771 and 81601577), (3) Post-doctoral Science Foundation of China (under Grant no.2016M592510), and (4) the Scientific Research Foundation for the Returned Overseas Chinese Scholars of school (LX2016N004).




IM/TH- Image Analysis (Single modality or Multi-modality): Machine learning

Contact Email