(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM
Room: Exhibit Hall | Forum 1
Purpose: In image guided radiation therapy (IGRT), cone-beam computed tomography (CBCT) system using kilo-voltage (KV) x-ray source is the most widely-used imaging devices. However, CBCT uses hazardous ionizing x-rays and hence minimizing the imaging dose is desirable based on ALARA principle. In this study, we propose a deep learning based non-iterative low-dose CBCT reconstruction method for IGRT applications.
Methods: In this study, we have formulated Low-dose CBCT reconstruction problem as restoring high quality high-dose CBCT projections (100kVp, 1.6 mAs) from noisy low-dose CBCT projections (100kVp, 0.1 mAs). The restoration of CBCT projection was performed using generative adversarial network (GAN) which is a convolutional neural net framework that consists of two models: 1) a generative model that produces high-dose CBCT projection from corresponding low-dose projection, and 2) a discriminative model that distinguishes with a certain probability if a given high-dose image is drawn from the true distribution of high-dose images or generated by the other network. Prior to the training, both images were filtered to eliminate low-frequency information which is not necessary for analytical reconstruction. For evaluation, the trained model was applied to unseen phantom data that was placed randomly on the couch. The restored high-dose projection was reconstructed using simple back-projection method.
Results: Training on 700 image pairs took approximately 16 hours to complete using Nvidia GTX 1080. PseudoCTs are produced by the trained model with a throughput time of approximately 80 projections/second. Significant noise reduction was achieved compared to original input while maintaining the quality comparable to the CBCT of high-dose projections.
Conclusion: The proposed deep learning-based method for CBCT reconstruction offers the ability to reducing the imaging dose without addition of reconstruction time. This makes our approach potentially useful in an on-line image-guided radiation therapy.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by Varian. J.S.K. acknowledges that this work was supported by Ministry of Science, ICT and Future Planning, Korea through the R&D program of NRF-2015M3A9E2067001.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
