A Probabilistic Approach to Optimizing Hardware Control Parameters in System Property Estimation using Gumbel-Softmax

Beomgu Kang, Hyunseok Seo

Keywords: categorical reparameterization, Gumbel Softmax, Optimization, System properties, physics model

Abstract:

Optimizing control parameters is crucial to estimate reliable tissue characteristics in quantitative MRI. Basically, multiple hardware parameters are simultaneously controlled to generate a signal from MRI system. Repetitive acquisitions with different control parameter combinations create distinct signal modulations and then tissue characteristics are deduced from prior knowledge of physics-based relationship among modulated signals, control parameters, and tissue characteristics. The choice of control parameters, which determines the attribute of signal modulation, directly impacts the inverse problem in tissue characteristic estimation. Thus, the multidimensional control parameter optimization remains an open research topic in MRI field for accurate analysis of tissue characteristics. Typically, optimal parameters are determined by iteratively updating sets of control parameters to maximize the estimation accuracy of the tissue characteristics. However, the conventional optimization process is restricted to explore only the vicinity of control parameters at the current iteration. Therefore, it could highly depend on initialization and current parameters, which might lead to inefficient search especially when noise is present in the system. In this work, to mitigate this limitation, we propose a novel Gumbel-Softmax-based optimization scheme that enables a probabilistic search across an expanding set of all candidates for each control parameter using categorical reparameterization. As a case study, the proposed method is employed to find optimal control parameters for quantitative MRI. We demonstrate that our Gumbel-Softmax-based optimization simultaneously explores the entire range of control parameters from early iterations and outperforms the conventional optimization approach on accuracy of MR tissue characteristic estimation and repeatability of optimization, especially under noisy environments.

Primary Area: optimization

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: 2036