DBGAN: A dual-branch generative adversarial network for undersampled MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Compressed Sensing Magnetic Resonance Imaging (CS-MRI) accelerates image acquisition but can compromise image quality.
• Deep learning, particularly Generative Adversarial Networks (GANs), has shown promise in improving CS-MRI reconstruction.
• Existing GAN-based methods using cascaded U-Net structures may suffer from limited feature propagation, leading to information loss.

Purpose of the Study:

• To propose a novel GAN-based model, DBGAN, for accelerated and high-quality MRI reconstruction from undersampled k-space data.
• To address information loss in cascaded architectures by enhancing feature propagation channels.
• To improve the efficiency and accuracy of MRI reconstruction for clinical applications.

Main Methods:

• Developed DBGAN, a GAN-based model featuring a generator with two end-to-end cascaded U-Nets.
• Implemented cross-stage skip connections (CSSC) to widen feature propagation channels between U-Net stages.
• Utilized instance normalization (IN) instead of batch normalization (BN) to mitigate training-inference discrepancies and incorporated a stage loss and a bilinear interpolation decoder branch.

Main Results:

• DBGAN achieved mean improvements of 3.65 dB in Peak Signal-to-Noise Ratio (PSNR) and 0.016 in Normalized Mean Square Error (NMSE) compared to state-of-the-art GAN methods on a T1-Weighted brain dataset.
• Qualitative results demonstrated considerable image reconstruction quality for both brain and knee MRI data across different modalities.
• The DBGAN model is computationally efficient, with less than half the parameters of existing methods and a reconstruction time of 60 milliseconds per 256x256 image.

Conclusions:

• DBGAN effectively reconstructs high-quality MRI images from undersampled k-space data, outperforming existing GAN-based methods.
• The proposed architectural enhancements, including CSSC and a bilinear interpolation decoder, significantly improve feature propagation and information recovery.
• DBGAN's efficiency and speed make it suitable for real-time MRI reconstruction, offering a promising advancement in medical imaging technology.