Exploring Siamese Networks in Self-Supervised Fast MRI Reconstruction

TL;DR

This paper introduces SiamRecon, a self-supervised Siamese network approach for MRI reconstruction that mimics an expectation maximization algorithm, achieving state-of-the-art accuracy without fully sampled data.

Contribution

The paper proposes a novel Siamese architecture for self-supervised MRI reconstruction, effectively avoiding trivial solutions and improving accuracy over existing methods.

Findings

Achieves state-of-the-art accuracy on brain and knee MRI datasets.

Mimics an expectation maximization algorithm for improved reconstruction.

Effectively avoids trivial solutions in self-supervised learning.

Abstract

Reconstructing MR images using deep neural networks from undersampled k-space data without using fully sampled training references offers significant value in practice, which is a self-supervised regression problem calling for effective prior knowledge and supervision. The Siamese architectures are motivated by the definition "invariance" and shows promising results in unsupervised visual representative learning. Building homologous transformed images and avoiding trivial solutions are two major challenges in Siamese-based self-supervised model. In this work, we explore Siamese architecture for MRI reconstruction in a self-supervised training fashion called SiamRecon. We show the proposed approach mimics an expectation maximization algorithm. The alternative optimization provide effective supervision signal and avoid collapse. The proposed SiamRecon achieves the state-of-the-art reconstruction accuracy in the field of self-supervised learning on both single-coil brain MRI and multi-coil knee MRI.