Let UNet Play an Adversarial Game: Investigating the Effect of Adversarial Training in Enhancing Low-Resolution MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Adversarial training is widely used for image realism but unexplored in clinical imaging.
• Magnetic Resonance Imaging (MRI) often suffers from low resolution and low signal-to-noise ratio (SNR).

Purpose of the Study:

• To investigate the efficacy of adversarial training for enhancing low-resolution and low-SNR MRI.
• To optimize adversarial and perceptual loss weights for clinical image enhancement.

Main Methods:

• Trained 206 networks on the OASIS-1 dataset with varying perceptual/adversarial loss weights and learning rates.
• Introduced Gradient Error (GE), Sharpness, and Edge-Contrast Error (ECE) metrics.
• Evaluated structural disparity, edge clarity, and pixel distribution distortion.

Main Results:

• Optimized adversarial loss reduced structural disparity (average SSIM reduction of 1.5%).
• Adversarial training significantly reduced GE (p < 0.05) and increased Sharpness.
• ECE reductions were observed for lower perceptual loss weights, with no significant adverse effects at higher weights.

Conclusions:

• Adversarial training significantly improves MRI enhancement pipelines.
• Highlights the need for hyperparameter optimization and novel image quality metrics in clinical imaging.