Improving Deep Learning MRI Reconstruction with Contrastive Learning Pretraining

Image reconstruction in Magnetic Resonance Imaging (MRI) aims to generate a high-quality image of the underlying anatomy from a minimal amount of measurements acquired during the scanning process. Deep learning (DL) has recently become the state-of-the-art approach for MRI reconstruction, often employing deep neural networks that learn to transform undersampled measurements into high-quality images, typically in a data-driven supervised manner. In our work, we propose a self-supervised contrastive learning pretraining framework for MRI that enhances the latent space for downstream reconstruction. Our framework is based on the fundamental principle of mutual information maximization, and we attempt to maximize the mutual information between different views of the same scan in latent space using contrastive learning. The learned latent space representations are subsequently utilized for reconstruction, and our experiments on various DL reconstruction models demonstrate significant improvements in performance both qualitatively and quantitatively.