Encrypted login | home

Program Information

A CNN Regression Approach for Volumetric Imaging Using Single X-Ray Projection

no image available
R Wei

R Wei1, F Zhou1 , B Liu1* , B Liang1 , Y Li1 , B Guo1 , X Xu1 ,S Xu2 (1) Image Processing Center, Beihang University, Beijing, Beijing;(2) PLA General Hospital, Beijing, China


SU-H4-GePD-J(B)-1 (Sunday, July 30, 2017) 4:30 PM - 5:00 PM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: Accurate volumetric imaging using breathing motion model and single X-ray projection is a useful technique for the management of respiratory motion. We developed and validated a new method based on convolutional neural network (CNN) to achieve this goal.

Methods: The method consisted of two stages: pre-treatment training and online application. In the training stage, a PCA based lung motion model with three principle components was first constructed using the planning 4D-CT. Then, by randomly sampling the PCA coefficients, 100 deformation vector fields (DVFs) were generated, and 100 training DRRs were calculated by projecting the deformed reference CT image using the generated DVFs. At last, a CNN model was trained which mapped the training DRRs to the corresponding PCA coefficients. In the application stage, the well trained CNN model was used to calculate the PCA coefficients and establish the 3D DVFs according to the real-time acquired X-ray projections. According to the generated DVFs, the volumetric image can be finally obtained.

Results: This method can calculate the 3D DVFs and volumetric image from single X-ray projection image. In a synthetic test, by randomly sampling the PCA coefficients, we generated 100 sets of test data, including ground truth volumetric images and test X-ray projection images. The average relative error of calculated PCA coefficients was less than 5%. The average normalized RMSE between the calculated and ground truth volumetric images was 0.0916%. A generalized gamma index was computed to quantify the differences from both spatial and intensity aspects. The pass rate was 100% for all tests, using gamma criteria of 3%/0.5 mm.

Conclusion: We proposed a volumetric imaging method based on CNN. This method could generate volumetric image from single X-ray image with high accuracy.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Natural Science Foundation of China (61601012).

Contact Email: