Leveraging Side Information with Deep Learning for Linear Inverse Problems: Applications to MR Image Reconstruction
Keywords: MR image reconstruction, side information, linear inverse problems, Trust-Guided Variational Network
TL;DR: We propose Trust-Guided Variational Network, a deep learning method that leverages side information to resolve ambiguities in linear inverse problems.
Abstract: Reducing the time it takes to acquire a Magnetic Resonance Imaging (MRI) scan is an important problem in healthcare, as it can improve patient care and reduce costs. One way to achieve this is by acquiring only a fraction of the frequency space data and reconstructing diagnostic-quality images from it. This problem can be formulated as a linear inverse problem (LIP), where the forward operator, which maps the structure of the imaged object to the acquired frequency space data, can become rank-deficient or exhibit many small singular values. This leads to ambiguities in the reconstruction process, where multiple images (most of them non-diagnostic) can map to the same set of acquired data. To resolve these ambiguities, it is essential to leverage domain knowledge and, whenever possible, exploit additional context (a.k.a., relevant side information) when solving the LIP. We present a novel, end-to-end trainable deep learning-based method, called Trust-Guided Variational Network (TGVN), that reliably incorporates side information into LIPs to eliminate undesirable solutions from the ambiguous space of the forward operator, while remaining faithful to the acquired data. We demonstrate its effectiveness through applications in multi-coil, multi-contrast MR image reconstruction, where incomplete or low-quality measurements from one contrast are used as side information to reconstruct a high-quality image of another contrast from heavily under-sampled data. Its robustness is validated by reconstructing images from different contrasts across different anatomies and field strengths. Compared to a set of baselines that also use side information, our method reconstructs high-quality images in the presence of heretofore challenging levels of under-sampling, thereby speeding up the acquisition drastically while providing protection against hallucinations. Our approach is also versatile enough to incorporate many different types of side information into any LIP.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 12651
Loading