(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 2
Purpose: MR guided radiation therapy (MRgRT) has become available lately due to technical advances. In MRgRT, MR images are acquired prior to a treatment to guide patient positioning and replanning if necessary. It is highly desirable to acquire the pre-treatment MR images at high speed. Undersampling data acquisition could speed-up data acquisition, but degrade image quality. One unique feature of MRgRT is the availability of high-quality MR images at treatment planning stage for the same patient. To obtain high-quality reconstructions for undersampled MR scans, we developed a deep manifold assisted algorithm using the planning MR image as a patient-specific prior.
Methods: Utilizing the high-quality planning MR image, we established a deep auto-encoder (DAE) to form a manifold of high-quality image patches, serving as the prior for reconstruction. The DAE was trained to map each input patch from the planning MR image onto a low-dimensional manifold, and then recover the input patch exactly. The trained manifold was then incorporated as a regularization to bring prior knowledge in reconstruction, restoring high-quality patches given input patches from undersampled reconstruction results. To demonstrate the effectiveness of the proposed algorithm, we performed a realistic simulation study using patient data. The high-quality prior MRI image was deformed to generate ground truth. We simulated undersampled MR data acquisition in a Cartesian grid with undersampling along the phase encoding direction. Noise was also introduced to simulate the effect of reducing signal average in data acquisition.
Results: We tested the proposed algorithm for undersampling ratios of 50%, 25%, 12.5%, and 6.25%. On average, the reconstruction error was reduced by 41.4%, while SSIM was improved by 20.5% compared to the conventional Fourier-transform based reconstruction method.
Conclusion: Employing the unique feature of the available high-quality patient-specific planning MR images, the proposed method achieved high-quality reconstruction for undersampled MR scans in MRgRT.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
