Exhibit Hall | Forum 2
Purpose: Gd-contrast-enhancement (CE) MRI is a powerful tool to visualize liver tumor. It is desired to use CE-MRI for accurate pretreatment tumor localization and target definition in adaptive replanning. However, repeated contrast injection over the treatment course raises a safety concern. This study develops a low-dose CE-MRI reconstruction method using deep learning and patient-specific pre-contrast MR image prior.
Methods: We trained a conditional Generative Adversarial Network (cGAN) to map a low-quality input CE-MR image to a high-quality one, based on the condition of a high-quality prior image acquired before contrast injection. The prior image and the image to be processed share similar image features but different CE structures and possibly deformed anatomy due to patient motions. We trained cGAN using liver patient cases each with a pre-contrast prior image and CE-MR images after contrast injection at different time points. Noise was added to post-contrast images to simulate cases with reduced signal-to-noise ratio. After the model was trained, we developed an iterative MRI reconstruction algorithm to incorporate the trained cGAN as a regularization to restore CE-MR images from k-space measurement data. We compared the performance of the proposed approach with that of the standard Fourier transform-based reconstruction method.
Results: In the case with a high contrast at 20 sec post injection without added noise, our method improved PSNR from 32.66 in the Fourier transform-based reconstruction method to 37.54. For the case with a low contrast level at 180 sec post injection and 15% added noise, our method improved PSNR from 21.07 to 25.36. Averaging all cases with different contrast levels and noise levels, our method increased PSNR by11.12dB, SNR by 39.83%, and SSIM by 25.13%.
Conclusion: Using patient-specific pre-contrast prior image can potentially reduce dose of contrast agent, facilitating low-dose CE-MRI in adaptive radiotherapy.