PIC-GAN: A Parallel Imaging Coupled Generative Adversarial Network for Accelerated Multi-Channel MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accelerated multi-channel MRI is crucial for reducing scan times and improving patient comfort.
• Existing reconstruction methods often struggle with noise and artifact reduction, especially at higher acceleration factors.
• Generative Adversarial Networks (GANs) show promise for image reconstruction tasks.

Purpose of the Study:

• To develop and evaluate a novel model, PIC-GAN, for accelerated multi-channel MRI reconstruction.
• To improve the quality of reconstructed MRI images by reducing noise and preserving structural details.
• To compare the performance of PIC-GAN against established reconstruction techniques.

Main Methods:

• Proposed a PIC-GAN model integrating parallel imaging (PI) with a GAN architecture.
• Incorporated data fidelity and regularization terms into the generator for end-to-end reconstruction.
• Combined adversarial and pixel-wise losses in image and frequency domains to preserve details.
• Evaluated on abdominal and knee MRI datasets with 2, 4, and 6-fold accelerations.

Main Results:

• PIC-GAN achieved the lowest Normalized Mean Square Error (NMSE) across abdominal and knee MRI datasets.
• PIC-GAN demonstrated the highest Peak Signal to Noise Ratio (PSNR) compared to L1-ESPIRiT, VN, and ZF-GAN.
• Experimental results showed effective reconstruction with low noise and improved structure similarity, outperforming other methods at 6-fold acceleration.

Conclusions:

• The PIC-GAN framework offers superior performance for accelerated multi-channel MRI reconstruction.
• PIC-GAN effectively reduces aliasing artifacts and restores tissue structures.
• The proposed method represents a significant advancement over conventional and state-of-the-art reconstruction techniques.