Click here to

Session: Multi-Disciplinary BLUE RIBBON [Return to Session]

Patient-Specific Image Prior Assisted Fast MR Imaging for Online Adaptive Radiotherapy

Y Gao*, C Shen, Y Gonzalez, J Deng, X Jia, University of Texas Southwestern Medical Center, Dallas, TX

Presentations

SU-I400-BReP-F2-3 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Magnetic resonance imaging (MRI) is increasingly used in radiotherapy to support pre-treatment image guidance and treatment adaptation. Fast MR imaging is of critical importance to ensure the validity of pre-treatment imaging guidance for accurate delivery and to facilitate a smooth clinical workflow. Undersampling and reduced data average can accelerate data acquisition, but lead to image artifacts and amplified noise. With the unique situation in radiotherapy with available high-quality patient-specific MR images for treatment planning, this study investigates the feasibility of using patient-specific MR images as prior to reconstruct MR images for fast imaging.

Methods: We trained a conditional Generative Adversarial Network (cGAN) to map a low-quality input MR image to a high-quality output one, based on the condition of input high-quality prior image of the same patient. The prior image and the image to be processed are from the same patient, sharing similar image features but deformed anatomy. We developed an iterative MR reconstruction algorithm to incorporate the trained network as a regularization to restore MR images from undersampled and noisy measurements. We trained cGAN using five patient cases each with 800 random noise signals added to the images. Data augmentation was performed by translating, rotating, and scaling the images. We compared the performance of the proposed approach with the same iterative MR reconstruction algorithm but with regularization trained with patient-generic image prior.

Results: In all the cases with 0-12.5% undersampling and 0-15% noise added, the proposed method with patient-specific prior consistently outperformed reconstruction methods with patient-generic prior. On average, our method improved PSNR by 5.4dB, SNR by 11.2%, and SSIM by 5.0%.

Conclusion: Taking advantage of patient-specific prior image in radiotherapy can potentially allow MR reconstruction with substantially undersampled and noisy data to facilitate fast imaging for online adaptive radiotherapy.

Keywords

MRI, Image Processing, Reconstruction

Taxonomy

IM- MRI : Reconstruction techniques

Contact Email

Share: