MRI super-resolution via realistic downsampling with adversarial learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Deep learning (DL) super-resolution (SR) models often perform poorly on real low-resolution (LR) magnetic resonance imaging (MR) data due to reliance on simulated LR images for training.
• The generalizability of SR networks is limited, hindering performance on actual clinical MR images.

Purpose of the Study:

• To develop a DL-based SR framework that improves performance on real LR MR images by emphasizing realistic data construction.
• To address the limitations of current SR methods in handling real-world MR image acquisition artifacts.

Main Methods:

• A two-step framework was proposed: (a) Downsampling training using a generative adversarial network (GAN) to create realistic LR/high-resolution (HR) pairs from real MR images.
• (b) Super-resolution training using the enhance4d deep super-resolution network (EDSR) on the GAN-generated data.
• Controlled experiments compared the proposed method against Gaussian blur and k-space zero-filling using liver MR images from 24 patients.

Main Results:

• The proposed method demonstrated significantly better resolution enhancement and fewer artifacts compared to Gaussian blur and k-space zero-filling.
• Outperformed the Gaussian method by 0.111 ± 0.016 in structural similarity index and 2.76 ± 0.98 dB in peak signal-to-noise ratio.
• Achieved a superior blind/reference-less image spatial quality evaluator metric (34.1 ± 2.4) compared to the Gaussian method (46.6 ± 4.2).

Conclusions:

• The proposed DL framework, with its focus on realistic data construction via GANs, effectively enhances SR performance on real LR MR images.
• This approach overcomes the limitations of conventional SR methods by generating more accurate training data, leading to superior image quality and reduced artifacts.
• The method shows significant potential for improving diagnostic accuracy in clinical MR imaging.