Room: Exhibit Hall | Forum 9
Purpose: Monte Carlo (MC) can accurately simulate the scatter in Cone-beam CT, but its clinical application is limited by the high computational time. In this work, we propose a novel simulation method which can estimate the scatter in CBCT effectively and accurately.
Methods: In order to cut the heavy simulation time of transporting particles, only 10 projections (512*384) of scatter sinogram are generated by MC. Then U-Net is applied to predict the remaining 350 projections. We trained our model using the head-neck dataset of 26 patients, in which each has 10 CBCT projections for inputs and 350 for labels. To avoid overfitting, data augmentation is employed, including patches cutting and shifting. The loss function of the model is mean square error (MSE), optimized with Adam solver. The method is implemented on GPU platform, using the Keras framework with the Tensorflow backend. To test the proposed method, we compare the simulation error and computational efficiency with the conventional MC method.
Results: After 7.0 hrs training time, the U-Net can be utilized to predict 350 projections from only 10 projections, testing within 0.5s, much faster than conventional MC method. In the testing case, mean absolute percentage error (MAPE) is lower than 2.97%, with the average value of 1.24%.
Conclusion: The proposed method indicates the potential of U-Net for fast scatter estimation. The achieved efficiency and accuracy make this method promising for MC-based scatter removal in CBCT.
Not Applicable / None Entered.