(Tuesday, 7/16/2019) 4:30 PM - 6:00 PM
Room: 225BCD
Purpose: To generate high quality Digitally Reconstructed Radiographs (DRR) from low resolution (LR) simulation CT volumes using a super resolution CNN. DRRs with improved quality will be useful for online image guidance.
Methods: A Convolutional Neural Network (CNN) was designed and trained to perform a 3D CT super-resolution (SR) from a low resolution (LR) volume to a high resolution (HR) volume with a 2x resolution. 10 publically available spinal datasets (SpineWeb) were used in the study. Out of the 10 datasets, 8 datasets were used to train the network, one dataset was used to valid and one dataset was used to test the network. The network was trained using image patches of size 32×32×32 that were randomly sampled from the training datasets. To maximize the bony structure sharpness in the predicted SR CTs and in the final DRRs, the network was trained to predict the difference images between the HR and LR images instead of directly predicting the HR images. The publically available DRR generation tool (Plastimatch) was then used to generate high quality DRRs from the network-predicted SR CT volumes.
Results: The proposed network was able to generate SR CT volumes from LR CT volumes. The DRRs generated using the generated SR images preserved much more details of the bony structures, e.g. edges, corners, spinal vertebras than the LR CT-generated counterparts.
Conclusion: Super resolution CNN is able to generate high quality DRRs with much improved anatomical structure definitions using low resolution CT volumes.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
