(Thursday, 8/2/2018) 1:00 PM - 3:00 PM
Room: Room 202
Purpose: X-ray scatter is the most severe contamination factor of cone-beam CT (CBCT), reducing its image quality and limiting clinical applications. The purpose of the study is to employ a Deep Residual Generative Adversarial Network (DRGAN) to remove scatter artifacts from a CBCT image.
Methods: A total number of 130 volumetric CT images were used in this study. We split data into a training set of 90 and a testing set of 40 images. Using an in-house GPU-based simulation package, we simulated CBCT projection data for each image. Signals with and without scatter were fed into FDK reconstruction algorithm to get the ground truth and scatter-contaminated CBCT images. We developed a DRGAN with Keras. The generator was a U-shape residual network with multiple 2D-convolutional, 2D-deconvolutional, and densely connected layers. It took a scatter-contaminated CBCT image slice as input and generated the scatter artifact to be removed. The discriminator was a deep 2D-convolutional network that mapped a CBCT image slice to a scalar indicating the probability of being ground truth. In addition to the traditional binary cross entropy loss, we added mean squared error (MSE) between the generator output and the corresponding ground truth to regularize training.
Results: After training 200 epochs, the network was able to remove the scatter artifacts for both training set and testing set. The relative MSE was reduced from ~4.5% to ~0.4% for the training set, and from ~5.1% to ~0.5% for the testing set. Average structural similarity was improved from ~0.86 to ~0.93 and from ~0.82 to ~0.91 for the two sets. Contrast to noise ratio was improved from ~7.0 to ~12.0.
Conclusion: We have built a Deep Residual GAN to remove X-ray scatter artifact from CBCT images. Image quality was significantly improved for both training and testing set.
Keywords
Not Applicable / None Entered.
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
