Room: Exhibit Hall | Forum 2
Purpose: MR-only guided radiation therapy (gRT) is an attractive goal for future gRT applications that takes advantage of the improved soft tissue contrast offered by MRI compared to CT. However, the electron density information used in dose calculations is derived from CT images, necessitating that CT simulation be a component of the conventional MR-gRT workflow. We propose a deep learning-based method for pseudoCT reconstruction that eliminates CT simulation from the MR-gRT workflow in a move towards MR-only gRT.
Methods: A generative adversarial network is a two-convolutional neural net framework that consists of a generator, which produces pseudoCT outputs based on MR inputs, and a discriminator, which differentiates between real and generated images. The networks are trained iteratively to improve the capabilities of each network. The framework was trained using paired CT and MR data from approximately 80 breast, lung, and abdomen patients previously treated at our institution. CT images were deformably registered to corresponding MR images. Both image sets were preprocessed to eliminate background elements outside the body contour and to create more homogeneous pixel ranges. For evaluation, the trained model was applied to unseen MR data and the generated pseudoCTs were used as the basis for dose calculations using the same plan parameters as clinically delivered plans.
Results: Training on 7000 image pairs took approximately 10 days to complete. PseudoCTs are produced by the trained model with a throughput time of approximately 0.1 s/slice. Dose volume histograms (DVHs) of the pseudoCT-based plans demonstrate target and organ at risk doses that are comparable at every DVH point to those of the clinically delivered plans.
Conclusion: The proposed deep learning-based method for pseudoCT reconstruction offers the potential to eliminate CT simulation from the MR-gRT workflow. A move towards MR-only gRT avoids registration related errors and extra radiation dose to the patient.
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.