Explanation-Driven Cyclic Learning for High-Quality Brain MRI Reconstruction from Unknown Degradation

Spatial resolution, signal-to-noise ratio (SNR), and motion artifacts critically matter in any Magnetic Resonance Imaging (MRI) practices. Unfortunately, it is difficult to achieve a trade-off between these factors. Scans with an increased spatial resolution require prolonged scan times and suffer from drastically reduced SNR. Increased scan time necessarily increases the potential of subject motion. Recently, end-to-end deep learning techniques have emerged as a post-acquisition method to deal with the above issues by reconstructing high-quality MRI images from various sources of degradation, such as motion, noise, and reduced resolution. However, those methods focus on a single known source of degradation, while multiple unknown sources of degradation commonly happen in a single scan. We aimed to develop a new methodology that enables high-quality MRI reconstruction from scans corrupted by a mixture of multiple unknown sources of degradation. We proposed a unified reconstruction framework based on explanation-driven cyclic learning. We designed an interpretation strategy for the neural networks, the Cross-Attention-Gradient (CAG), which generates pixel-level explanations from degraded images to enhance reconstruction with degradation-specific knowledge. We developed a cyclic learning scheme that comprises a front-end classification task and a back-end image reconstruction task, circularly shares knowledge between different tasks and benefits from multi-task learning. We assessed our method on three public datasets, including the real and clean MRI scans from 140 subjects with simulated degradation, and the real and motion-degraded MRI scans from 10 subjects. We identified 5 sources of degradation for the simulated data. Experimental results demonstrated that our approach achieved superior reconstructions in motion correction, SNR improvement, and resolution enhancement, as compared to state-of-the-art methods.