Compressed sensing using a deep adaptive perceptual generative adversarial network for MRI reconstruction from undersampled K-space data

Objective: Magnetic resonance imaging (MRI) reconstruction from undersampled k-space data has received great interest due to its capability in reducing physical scan time. Meanwhile, the reconstruction problem is challenging because of its ill-posed and inverse nature. Nevertheless, existing compressed sensing (CS) and deep learning-based methods still need improvement as they suffer from limited generalisation ability, especially when higher undersampling factors are applied.

Methods: To recover high-quality images with reliable fine anatomical structures, we propose DAPGAN - a deep adaptive perceptual generative adversarial network which reconstructs high-quality MR images from undersampled k-space data. In particular, a novel perceptual feature guidance (PFG) mechanism is proposed which has the capability to retrieve effective features from each level that is useful in emphasising underlying anatomical structures. In addition, the model explores information in a dual-domain style.

Results: Experimental results show that the proposed method outperforms state-of-the-art baselines in terms of quantitative and qualitative evaluations. Our method improves the average SSIM (structure similarity index measurement) from 0.81 to 0.93 at a low CS ratio of 10%, compared to the average performance of competing methods on cardiac datasets using Cartesian sampling.

Conclusion: An innovative mechanism was proposed for accurate and perceptual feature guidance. It is an adaptive error-correction-based mechanism during multi-level feature reconstruction. The effectiveness was proved by the superior performance in extracting reliable anatomical details.

Significance: The architecture of our proposed model offers a new solution for accurate feature guidance, considering enhancing conventional optimisation-based problems. In particular, it is a robust mechanism in aggressive undersampling scenarios.