Real Time Volumetric MRI From Ultra-Sparse Samples Via Implicit Neural Representation Learning
R Yesiloglu1*, L Shen2, A Johansson3, Y Cao4, J Balter5, L Xing6, L Liu7, (1) Stanford University, Stanford, CA, (2)Stanford University,Palo Alto, CA, (3) Uppsala University, Sweden, (4) University of Michigan, Ann Arbor, MI, (5) University of Michigan, Ann Arbor, MI, (6) Stanford University School of Medicine, Stanford, CA, (7) Stanford University, Palo Alto, CA
Presentations
MO-E115-IePD-F6-5 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 6
Purpose: To investigate the feasibility of real time volumetric MRI with ultra-sparse sampling by learning implicit neural representations of image intensities given a single fully-sampled volume as the reference.
Methods: Eight abdominal patients with free-breathing volumetric MRI time series (5 min scan length and 800 image volumes) were included. For each patient, image volume at the first time point was selected as the fully-sampled reference and we reconstructed the remaining volumes of different motion states from sparse samples of 2 radial projections only with an acquisition time of 340 ms. An implicit neural representation network with two cascaded blocks was investigated. The first block learns a transformation between volumetric coordinates of the reference and the target volumes, i.e. the deformation field between the two and the second block learns a mapping between transformed volumetric coordinates of the target and the corresponding voxel intensities. The network was first trained using the reference volume by minimizing the difference between the network estimated and ground truth voxel intensities. To reconstruct target volumes with sparse sampling, the network weights were updated by minimizing the network estimated and ground truth voxel intensities at sampled locations with a Jacobian regularizer on the estimated deformation field. We evaluated the image reconstruction quality by calculating PSNR and SSIM between reconstructions and the ground truth.
Results: Across all the patients evaluated, the PSNR between network reconstructions and the ground truth ranged between 34.93 dB and 45.46 dB with an average of 41.97 dB. The SSIM ranged between 0.945 and 0.995 with an average of 0.987.
Conclusion: Quantitative evaluation of image quality metrics suggested the proposed implicit neural representation network can reconstruct high quality volumetric MRI with sub-second sampling time, which has the potential of supporting real time imaging guidance during radiotherapy such as 3D motion tracking during treatment delivery.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01 EB016079 and Stanford Institute of Human-Centered Artificial Intelligence. The radial scanning sequence was provided under a research agreement from Siemens Healthineers.
Keywords
MRI, Reconstruction, Image-guided Therapy
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking
Contact Email