Room: Room 202
Purpose: Cone-beam computed tomography (CBCT) is widely used in the clinic worldwide. However, its image quality is severely degraded by scatter. This study proposed a nearly-real-time scatter correction method using a residual convolutional neural network (CNN) in projection domain.
Methods: An Unet-based CNN with 19 convolutional layers was generated to estimate scatter intensity in CBCT projections. Monte Carlo simulation was performed to collect 900 projections of CIRS, sphere and head phantoms for training of the deep learning-based scatter correction model. For evaluation of the model, we collected 360 Gammex phantom projections. We set inputs and labels of the CNN to measured projections and scatter-only projections respectively. Data augmentations of small patch extraction, flip and 90-degree rotations were randomly applied for the training dataset. Total 200-epoch end-to-end training was implemented with mini-batch size of 10. CNN-corrected projections and their reconstructed Hounsfield Unit (HU) images were compared with projections and images corrected using a conventional adaptive scatter kernel superposition (ASKS)-based method.
Results: The CNN-corrected projections accurately represented scatter-free projections. Averaged peak signal-to-noise ratio and its standard deviation over reconstructed Gammex phantom images were improved from 8.54 Â± 4.55 [dB] for uncorrected images to 12.0 Â± 4.9 [dB] for the CNN-corrected images in contrast to 10.3 Â± 4.8 [dB] for the ASKS-corrected images. The CNN-corrected reconstructed images had better HU intensity accuracy than the ASKS-corrected reconstructed images. Computation time for calculating 360 projections was around 2.9 seconds whereas the same calculation required around 5.5 minutes for ASKS correction.
Conclusion: A novel projection-based CBCT scatter correction method using CNN was developed. The CNN-correction offers more accurate corrected projections with much less time-cost than the conventional ASKS-based correction.