Click here to

Session: Multi-Disciplinary: Image Guidance: Cone-Beam CT [Return to Session]

Synthetic Contrast MR Image Generation Using Deep Learning

Y Liu*, T Wang, Y Lei, J Roper, J Bradley, T Liu, X Yang, Emory Univ, Atlanta, GA


SU-IePD-TRACK 3-6 (Sunday, 7/25/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: Gadolinium-based dye is often used as the contrast agent during MR imaging examinations. In addition to extra cost, labor and time, contrast-enhanced MR imaging can cause patient discomfort and side effects. We propose a deep learning-based method to synthesize contrast-enhanced MR from non-contrast MR images.

Methods: A cohort of 274 brain cancer patients from a public dataset, the Brain Tumor Segmentation Challenge (BRATS2015) was used. They have both non-contrast and contrast-enhanced MR images. A self-attention cycle generative adversarial network (cycleGAN) was used to learn the mapping from non-contrast images to contrast MR images. Attention U-Net was used as a generator of CycleGAN to force the trained model to extract information that can represent the specific difference between non-contrast MRI and contrast MRI, namely, the tissue contrast difference. The proposed method was trained on 224 patients and evaluated on the rest 50 patients.

Results: The discrepancy between the ground truth contrast MR and synthetic contrast MR images were compared. The mean absolute error (NMAE) was 0.035±0.004. The peak signal-to-noise ratio (PSNR) was 32.4±2.6. And structural similarity index (SSIM) was 0.97±0.06.

Conclusion: It is shown that the proposed method has great potential in accurately generating contrast MR, and therefore bypassing the contrast administration step during MR scan, which can save labor, time, cost and mitigate patient discomfort.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email