Room: Exhibit Hall | Forum 5
Purpose: Shading artifacts due to scatter contamination degrade conebeam CT (CBCT) image quality. We proposed an image-domain shading correction scheme recently by extracting the compensation map and adding it to the uncorrected image. An accurate compensation map is not easy to acquire using conventional segmentation on CBCT image with severe shading artifacts. Inspired by the self-learning capability and simplicity of the convolutional neural network (CNN) and the residual learning framework, we propose a novel scatter correction scheme for CBCT using deep residual CNN network.
Methods: We add a shortcut connection based on U-net-like network between input and output to set up the residual CNN for training the compensation map with reference to the layer inputs. The residual network facilitates the construction of deep CNN since it mitigates the gradient vanishing problem. Other shortcut connections are designed to and incorporate low and high level features extracted by convolution, pooling and upsampling operations. In this work, Monte Carlo simulation is chosen to generate the label images and the ground truth for comparison. We use 1954 slices of the pelvis images from 11 patients scanned by Varian Trilogy OBI system as inputs. Training data and validation data account for 70% and 30% of the inputs, respectively. Two more patient data are predicted using the trained model for reproducibility test.
Results: Compared with the ground truth, the RMSE of images is reduced from 144 to 11 HU and the SU is increased by 20% after scatter correction. The structural similarity is slightly increased from 0.95 to 0.99, indicating that the scatter correction maintains the anatomical structure of the patient.
Conclusion: We design the deep residual CNN incorporating a U-net-like network with shortcut connections to train the compensation map. This method achieves high scatter correction efficacy with the advantage of robustness of batch data correction.