Click here to

Session: Science Council Session: Innovative Technologies to Advance Diagnosis and Treatment [Return to Session]

Patient-Specific Deep Learning-Based Self-High-Resolution for MR Imaging

Y Lei*, J Roper, E Schreibmann, H Mao, J Bradley, T Liu, X Yang, Emory Univ, Atlanta, GA


TU-EF-TRACK 4-6 (Tuesday, 7/27/2021) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: High-resolution (HR) magnetic resonance imaging (MRI) is desirable in many clinical applications, but there is a trade-off between image resolution and acquisition time. HR generation methods may yield fine details from fast acquisitions; however, most current methods attempt to solve this ill-posed problem using external atlases to derive the transformation from low-resolution (LR) to HR images. In this study, we propose a deep-learning-based self-HR method, which instead generates HR images directly from a patient’s thick-cut MRI.

Methods: We propose a new deep-learning-based, patient-specific method using a self-HR framework. Two LR-to-HR deep-learning models are pre-trained on HR in-plane images from a training dataset and then are fine-tuned on each test patient’s in-plane images. To generate HR through-plane MR images for a test patient, the LR sagittal and coronal images are input into the trained model, and the resulting HR MRIs are fused together to obtain a HR 3D MRI. This technique was validated across 50 clinical datasets of multimodal MRIs: Flair, T1-weighted contrast and T2-weighted images. The native HR images were used as the ground truth for evaluating resolution recovery from down-sampled images.

Results: The normalized mean absolute error (NMAE) and peak-signal-to-noise-ratio (PSNR) were calculated between the generated and original HR images. The LR data was obtained via down-sampling the MRI in z-axis. The down-sampling factor was set to 3. Overall, the mean NMAE and PSNR were 0.019±0.005 and 32.1±1.6 dB for Flair, 0.018±0.005 and 32.4±1.5 dB for T1c, and 0.015±0.003 and 33.7±1.1 dB for T2, which demonstrated the accuracy of the proposed method.

Conclusion: We have proposed a patient-specific deep learning-based self-HR method to derive high through-plane resolution MR images and validated its feasibility. This proof-of-concept study shows potential to improve image resolution without any high through-plane resolution MR data and its self-HR capability could have broad clinical applications.



    High-resolution Imaging, MRI


    IM- MRI : Machine learning, computer vision

    Contact Email