Self-supervision Meets Bootstrap Estimation: New Paradigm for Unsupervised Reconstruction with Uncertainty Quantification

Zhengliang Wu, Longbing Cao, Xuesong Li

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: self-supervised learning, MRI reconstruction, uncertainty quantification, compressed sensing

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: The paper presents a novel methodology, yielding advanced and interpretable results in the context of self-supervised MRI reconstruction, by connecting self-supervision with Bootstrap Estimation.

Abstract: Deep learning-based self-supervised reconstruction (SSR) plays a vital role in diverse domains, including unsupervisedly reconstructing magnetic resonance imaging (MRI). Current powerful methodologies for self-supervised MRI reconstruction usually rely on capturing the relationships between different views or transformations of the same data such as serving as inputs and labels respectively, which show notable influence from analogous approaches in computer vision. Although yielding somewhat promising results, their designs are often heuristic without deep insights into reconstructed object characteristics, and the analytical and mathematical principles of such methods are not expressive. This paper addresses these issues by a novel SSR paradigm, BootRec, that not only provides a theoretical foundation for self-supervised reconstruction but also facilitates the development of downstream algorithms. Self-supervised MRI reconstruction is modeled as error-oriented parameter estimation - Bootstrap estimation for SSR (BootRec). In BootRec, we demonstrate the mathematical equivalence between bootstrapping in a sample set and the commonly used re-undersampling operation for SSR. This insight is further incorporated into designing models to estimate the variances and errors of MRI SSR results without accessing labeled data. The error estimation serves as the loss function for unsupervisedly training the models. Empirical experiments show that our new paradigm BootRec enables effective uncertainty quantification and advanced MRI reconstruction performance against other zero-shot methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7561