Click here to


Are you sure ?

Yes, do it No, cancel

Transfer Learning of a Convolutional Neural Network for CBCT Projection-Domain Scatter Correction with Different Scan Conditions

Y Nomura1*, Q Xu2,3 , H Shirato2,4 , S Shimizu2,5 , L Xing2,6 , (1) Department of Radiation Oncology, Graduate School of Medicine, Hokkaido University, Sapporo, Japan, (2) Global Station for Quantum Medical Science and Engineering, Global Institution for Collaborative Research and Education (GI-CoRE), Hokkaido University, Sapporo, Japan, (3) Beijing Engineering Research Center of Radiographic Techniques and Equipment, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing, China, (4) Department of Radiation Medicine, Faculty of Medicine and Graduate School of Medicine, Hokkaido University, Sapporo, Japan, (5) Department of Radiation Medical Science and Engineering, Faculty of Medicine, Hokkaido University, Sapporo, Japan, (6) Department of Radiation Oncology, Stanford University, Stanford, CA


(Tuesday, 7/16/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 9

Purpose: Convolutional neural network (CNN) is a promising method for cone beam CT (CBCT) scatter correction in the projection domain. However, a trained CNN model corrects the scatter only for a specific scan condition and needs to be trained from scratch for different conditions. This study aims to evaluate feasibility of transfer learning for CNN-based scatter correction with a small number of additional training datasets when the scanning condition is changed from full-fan to half-fan.

Methods: An existing trained CNN model for full-fan projections was tuned to extract scatter distributions from measured half-fan projections with transfer learning. Monte Carlo simulation was performed to collect in total 360 half-fan projections of six non-anthropomorphic phantoms for fine-tuning. For evaluation, 360 projections of an anthropomorphic lung phantom were collected. Fifteen-epoch end-to-end training was implemented only for parameters of the last two convolutional layers and last batch normalisation layer, while others remained unchanged. The accuracy of the CNN model with tuning was compared with the CNN model without tuning and a hand-crafted method called fast adaptive scatter kernel superposition (fASKS).

Results: The tuned-CNN model provides statistically significant image quality improvements compared to the fASKS method or non-tuned-CNN model. Averaged structural similarity and its standard deviation of lung phantom images are improved from 0.9982 ± 0.0008 for the non-tuned CNN correction to 0.9992 ± 0.0002 for the tuned CNN correction in contrast to 0.9990 ± 0.0003 for the fASKS correction. Computation time of the tuned CNN correction for calculating 360 projections is around 4.8 seconds whereas it requires around 5.5 minutes for the fASKS correction.

Conclusion: This study demonstrates that the trained CNN model is expandable for different scan conditions with a small number of additional training datasets. The tuned CNN model provides more accurate scatter-corrected projections than the fASKS correction and non-tuned CNN model.


Cone-beam CT, Scatter, Quantitative Imaging


IM- Cone Beam CT: Machine learning, computer vision

Contact Email