Skip to content

Efficient Noise Calculation in Deep Learning-based MRI Reconstructions

Conference: ICML2025
arXiv: 2505.02007
Code: To be released
Area: Medical Imaging / Uncertainty Quantification
Keywords: MRI reconstruction, noise propagation, Jacobian Sketching, voxel variance, uncertainty quantification, g-factor

TL;DR

An efficient method based on Jacobian Sketching is proposed. By probing the Jacobian diagonal elements of DL reconstruction networks via random phase vectors, it accelerates the computation of voxel-level noise variance in MRI reconstruction using an unbiased estimator. The computation and memory requirements are reduced by more than an order of magnitude while maintaining a 99.8% correlation coefficient with the Monte Carlo reference.

Background & Motivation

Background

Background: Noise analysis in classical pMRI: Classical parallel imaging methods like SENSE have explicit g-factor analyses to evaluate spatial noise amplification.

Limitations of Prior Work

Limitations of Prior Work: Lack of noise analysis in DL reconstruction: Due to the non-linear complexity of neural networks, existing DL reconstruction methods typically ignore noise propagation and rely solely on global metrics such as PSNR/SSIM.

Key Challenge

Key Challenge: Limitations of Monte Carlo: Traditional MC methods require thousands of repeated reconstructions, which is computationally prohibitive, especially for 3D/4D data.

Core Idea

Core Idea: Importance of noise analysis: It directly impacts SNR evaluation, sampling strategy design, network architecture selection, and confidence in clinical deployment.

Supplementary Notes

Supplementary Notes: Threefold contributions: (1) A theoretical framework connecting k-space noise with image variance; (2) An efficient implementation of Jacobian Sketching; (3) Extensive validation across different architectures and training paradigms.

Method

Core Theory

  • Noise Covariance: For a reconstruction function \(f\), after first-order Taylor expansion: $\(\bm{\Sigma}_{\bm{x}} = \bm{J}_f \bm{A}^H \bm{\Sigma}_k \bm{A} \bm{J}_f^H = \bm{L}\bm{L}^H\)$ where \(\bm{L} = \bm{J}_f \bm{A}^H \bm{\sigma}_k\), and \(\bm{J}_f\) is the network Jacobian.
  • Unbiased Estimator (Theorem 3.1): For any random vector \(\bm{v}\) satisfying \(\mathbb{E}[\bm{v}\bm{v}^H]=\bm{I}\) : $\(\mathbb{E}[(\bm{\Sigma}\bm{v}) \odot \bm{v}^*] = \text{diag}(\bm{\Sigma})\)$
  • Cholesky Decomposition Simplification (Lemma 3.2): "Voxel variance = \(\|\bm{l}_i\|_2^2\)", eliminating the need to explicitly compute the full Jacobian.

Jacobian Sketching Algorithm

  1. Generate a random phase vector matrix \(\bm{V}_S \in \mathbb{C}^{m \times S}\) (\(v_i = e^{j\theta_i}\)).
  2. Transform via \(\bm{\sigma}_k\) and \(\bm{A}^H\): \(\widetilde{\bm{W}}_S = \bm{A}^H \bm{\sigma}_k \bm{V}_S\).
  3. Compute \(\bm{U}_S = \bm{J}_f \widetilde{\bm{W}}_S\) via JVP.
  4. Hadamard product + averaging: \(\widehat{\text{diag}(\bm{\Sigma}_{\bm{x}})} = \frac{1}{S} \bm{U}_S \odot \bm{U}_S^H \cdot \mathbf{1}_S\).

Random Vector Selection

  • Complex Rademacher (random-phase) vectors have lower variance than standard complex Gaussians, making them more suitable for MRI scenarios.
  • \(S=1000\) probing vectors are sufficient to achieve high accuracy.

Key Experimental Results

Main Results: Generalization Across Architectures (R=8, α=1)

Method Knee PCC(%) Knee NRMSE(%) Brain PCC(%) Brain NRMSE(%)
E2E-VarNet 99.9±0.0 0.7±0.0 99.9±0.0 0.5±0.1
MoDL 99.9±0.0 0.5±0.0 99.7±0.0 1.1±0.1
U-Net 99.4±0.0 1.7±0.2 99.7±0.0 1.8±0.2
SSDU 99.9±0.0 - - -

The average correlation coefficient across all methods is 99.8%, with an average error of 0.8%.

Ablation Study

  • Robustness to Noise Levels: Maintains high accuracy within the range of \(\alpha \in \{1, 5, 10, ..., 200\}\).
  • Robustness to Acceleration Factors: Valid under R=4, 8, and 12.
  • Robustness to Sampling Patterns: Stable under various undersampling patterns such as Poisson Disc and random sampling.
  • Computational Efficiency: Computation is approximately 1/10 to 1/100 of MC with 3,000 iterations.

Highlights & Insights

  1. Rigorous yet Practical Theory: Complete logical chain from first-order approximation \(\rightarrow\) covariance decomposition \(\rightarrow\) unbiased estimator \(\rightarrow\) efficient implementation.
  2. Architecture-Agnostic: Only requires the network to be differentiable (existence of Jacobian), making it suitable for both data-driven and physics-driven architectures.
  3. Filling the Gap in DL-MRI: Reintroduces noise analysis as a core component of reconstruction algorithms.
  4. Complex Rademacher Vectors: A customized random probing scheme tailored for MRI scenarios, offering lower variance.
  5. Clinical Significance: Variance maps can directly guide radiologists to identify regions of noise amplification, enhancing diagnostic confidence.
  6. Value in Unsupervised Scenarios: In unsupervised training regimes like SSDU where fully-sampled reference data is unavailable, SNR maps can replace PSNR/SSIM for quality assessment.

Relationship with Traditional g-factor

  • Classical SENSE g-factor: \(g_i = \sqrt{[(\bm{S}^H\bm{\Psi}^{-1}\bm{S})^{-1}]_{ii} \cdot [\bm{S}^H\bm{\Psi}^{-1}\bm{S}]_{ii}}\)
  • The proposed method can be viewed as a natural generalization of the g-factor in non-linear DL reconstruction, preserving voxel-level spatial resolution.
  • When \(f\) is linear, the proposed method degenerates exactly to the classical g-factor analysis.

Limitations & Future Work

  • Modeled based on first-order Taylor approximation, which may introduce errors in highly non-linear networks (such as pure U-Net).
  • Propagation of noise in generative models (e.g., diffusion models) is not considered, where stochasticity introduces additional complexity.
  • Only handles acquisition noise (aleatoric uncertainty), leaving epistemic uncertainty uncovered.
  • Linear approximation may fail in extremely low SNR regions.
  • Practical application on 3D volumetric data is not yet explored (only 2D slices have been validated).
  • Integrating variance maps with clinical workflows to show noise amplification areas in real-time remains future work.
  • Classical g-factor (Pruessmann et al., 1999): The proposed method serves as its natural generalization in DL reconstruction.
  • Hutchinson Estimator (1990): The foundation of diagonal estimation for real-valued matrices, which this work extends to the complex domain and integrates with MRI operators.
  • Dawood et al. (2024): Analytical noise estimation designed specifically for k-space interpolation networks; the proposed method is more general.
  • SSDU (Yaman et al., 2020), N2R (Desai et al., 2022): Unsupervised/semi-supervised training paradigms, both of which are validated to be compatible with this work.
  • Insights: Can be utilized to guide noise-aware sampling strategy design, neural architecture search, and real-time integration with clinical workflows.

Rating

  • Novelty: ⭐⭐⭐⭐
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐
  • Writing Quality: ⭐⭐⭐⭐⭐
  • Value: ⭐⭐⭐⭐⭐