Implicit Regression in Subspace for High-Sensitivity CEST Imaging

Chu Chen, Yang Liu, Se Weon Park, Jizhou Li, Kannie W. Y. Chan, Raymond H. F. ChanISBI 2024

Abstract

Chemical Exchange Saturation Transfer (CEST) MRI is a powerful molecular imaging technique capable of detecting low-concentration proteins and metabolites. However, its clinical utility is often hampered by low signal-to-noise ratios (SNR) and the resulting artifacts in z-spectra. This paper presents IRIS (Implicit Regression in Subspace), an unsupervised denoising framework that integrates subspace decomposition with coordinate-based neural networks. By leveraging the inherent redundancy in CEST data and the spectral bias of neural networks, IRIS effectively recovers high-fidelity signals without the need for labeled training datasets. Experimental results on synthetic phantoms and in-vivo animal models demonstrate that IRIS significantly outperforms traditional and supervised deep-learning methods in both quantitative metrics and visual clarity.

Introduction

CEST MRI provides unique insights into biological processes by "borrowing" signal from the abundant water pool to detect dilute solutes like amides and amines. This makes it particularly valuable for identifying brain tumors and other metabolic abnormalities. Despite its potential, the resulting z-spectra are notoriously noisy. Existing denoising strategies typically fall into two categories: traditional algorithms like Singular Value Decomposition (SVD), which often suffer from over-smoothing or residual noise, and modern supervised deep-learning models like U-Nets, which require massive amounts of high-quality ground-truth data—a resource that is rarely available in medical imaging.

To bridge this gap, we propose IRIS, a method that operates in a data-efficient, unsupervised manner. IRIS treats denoising as a regression problem within a low-dimensional subspace, utilizing the inductive bias of neural networks to separate signal from noise.

Methodology

The IRIS framework is built upon two foundational pillars: subspace formulation and implicit neural representation.

Subspace Formulation

CEST data is highly redundant because the spectral shapes across different pixels share common underlying physical characteristics. IRIS exploits this by using SVD to decompose the noisy data into a low-dimensional subspace. This process separates the data into a "basis" (representing the general spectral shapes) and "spatial coefficients" (representing the weight of those shapes at each pixel). By restricting the reconstruction to a small number of basis functions ($K$), the complexity of the denoising task is significantly reduced.

Implicit Neural Representation (INR)

The core of the "Implicit Regression" involves a lightweight Multi-Layer Perceptron (MLP) referred to as RegressionNet. Instead of processing the entire image at once, the network takes pixel coordinates $(i, j)$ as input and learns to predict the corresponding spatial coefficients.

Because neural networks are naturally biased toward learning smooth, continuous functions (a phenomenon known as spectral bias), the MLP effectively fits the underlying anatomical signal while ignoring the high-frequency, stochastic noise. This allows the model to be trained directly on the noisy scan itself, eliminating the need for external labels or clean ground-truth images.

Experimental Setup

The performance of IRIS was validated using two distinct datasets:

  1. Synthetic Phantom: A simulated dataset based on the Lorentzian model. It featured various shapes (square, circular, octagonal) to represent different chemical effects, including direct water saturation, relayed Nuclear Overhauser Effects (rNOE), and Amide Proton Transfer (APT).
  2. In-vivo Animal Experiment: Data from mice injected with U-87 MG tumor cells. This provided a complex biological environment to test the detection of APT and rNOE signals in a clinical-analogous setting.

Results and Discussion

Quantitative Performance

In comparative testing under severe noise conditions, IRIS demonstrated substantial improvements in Peak Signal-to-Noise Ratio (PSNR). IRIS achieved a PSNR of 27.87 dB, whereas traditional MLSVD reached only 20.91 dB. Notably, the supervised deep-learning method DECENT performed poorly in these specific conditions, trailing at 16.10 dB. This highlights the robustness of the IRIS unsupervised approach when high-quality training data is unavailable.

Qualitative Visual Fidelity

Qualitatively, IRIS produced the clearest tumor maps. While competing methods often introduced "speckle" noise or blurred the boundaries between the tumor and healthy tissue, IRIS maintained sharp edges and high contrast. This clarity is essential for the accurate localization of metabolic markers like APT and rNOE in oncological applications.

Computational Efficiency

By utilizing a lightweight MLP rather than a heavy convolutional architecture, IRIS remains computationally manageable. However, the iterative nature of the optimization—requiring approximately 3,000 iterations per slice—remains a point for future optimization.

Conclusion

IRIS represents a significant advancement in CEST MRI denoising by combining the mathematical rigor of subspace modeling with the flexible regularizing power of implicit neural representations. Its ability to perform high-sensitivity imaging without labeled data makes it a promising candidate for clinical translation. Future work will focus on accelerating the convergence speed and validating the framework on diverse human clinical datasets across various scanner strengths.

Summary updated

BibTeX
@inproceedings{chen2024implicit,
  author = {Chu Chen and Yang Liu and Se Weon Park and Jizhou Li and Kannie W. Y. Chan and Raymond H. F. Chan},
  title = {{Implicit Regression in Subspace for High-Sensitivity CEST Imaging}},
  booktitle = {ISBI},
  year = {2024},
}
Figure