Rician likelihood loss for quantitative MRI using self-supervised deep learning

Christopher S. Parker, Anna Schroder, Sean C. Epstein, James Cole, Daniel C. Alexander, Hui Zhang

Introduction

Quantitative MRI (qMRI) aims to transform qualitative image intensities into maps of physical tissue parameters, such as diffusion coefficients. While traditional voxel-wise curve fitting is computationally intensive, deep learning has revolutionized the field by providing near-instantaneous parameter estimation. Self-supervised learning is particularly attractive for qMRI, as it allows networks to learn from raw data by minimizing the difference between measured signals and those predicted by a biophysical model.

The Challenge: Rician Noise and the Gaussian Assumption

Most self-supervised qMRI networks utilize a Mean Squared Error (MSE) loss function. MSE implicitly assumes that noise follows a zero-mean Gaussian distribution. However, MRI magnitude signals follow a Rician distribution. At high Signal-to-Noise Ratios (SNR), the Rician distribution approximates a Gaussian. But at low SNR, the Rician distribution becomes skewed, and its expected value is higher than the true underlying signal. This creates a "noise floor" that systematically biases measurements.

The Rician bias: At low SNR, the expected value of the measured signal (dashed line) deviates significantly from the true signal A (solid line) due to the noise floor.

When a network trained with MSE attempts to fit noisy, low-SNR data, it mistakenly interprets this elevated noise floor as part of the actual signal. This leads to systematic errors, such as the underestimation of diffusion coefficients in regions with low signal intensity.

Methods: A Numerically Stable Rician Loss

To address this, researchers developed a Negative Log Rician (NLR) likelihood loss function. While the Rician likelihood is theoretically superior for magnitude MRI, its implementation in deep learning has been hindered by numerical instability. The formula involves modified Bessel functions ($I_0$) that grow exponentially, often causing "overflow" errors during gradient descent.

The authors overcame this by implementing a numerically stable version using:

  1. Exponentially scaled Bessel functions ($I_0^e$): These keep values within a manageable range for standard computational hardware.
  2. Chebyshev polynomials: These provide high-accuracy approximations of the log-Bessel function across the entire SNR spectrum.

This NLR loss was evaluated using two common biophysical models: the simple Apparent Diffusion Coefficient (ADC) model and the more complex Intra-Voxel Incoherent Motion (IVIM) model, which separates tissue diffusion from blood microcirculation.

Results and Performance

The proposed NLR loss was tested against standard MSE using simulated datasets (200,000 voxels) where ground truth parameters were known. The results demonstrate that the NLR loss effectively eliminates the systematic bias inherent in MSE-based training without sacrificing precision.

Performance comparison: The NLR loss (blue) maintains near-zero bias across parameter ranges for both ADC and IVIM models, whereas the MSE loss (orange) shows significant systematic bias as SNR decreases.

Key Findings:

  • Bias Elimination: At an SNR of 10, MSE loss resulted in a 5% bias for ADC and a substantial 20% bias for IVIM diffusion coefficients. In contrast, the NLR loss reduced this bias to approximately 1%.
  • High SNR Robustness: In high-SNR scenarios where Rician noise is approximately Gaussian, the NLR loss performed identically to MSE, making it a safe "plug-and-play" replacement.
  • Numerical Stability: The implementation remained stable across a wide SNR range (0–40), outperforming previous approximations that often failed at specific signal levels.

Conclusion

By replacing the standard MSE with a numerically stable Rician likelihood loss, self-supervised deep learning models can achieve high-accuracy parameter mapping even in low-SNR clinical conditions. This approach corrects the systematic underestimation of tissue parameters caused by the Rician noise floor, providing more reliable quantitative maps for disease detection and monitoring. Future work may involve extending this loss function to other qMRI techniques like $T_1/T_2$ relaxometry and incorporating Bayesian priors to further enhance estimation accuracy.

Summary updated

BibTeX
@inproceedings{parker2025rician,
  author = {Christopher S. Parker and Anna Schroder and Sean C. Epstein and James Cole and Daniel C. Alexander and Hui Zhang},
  title = {{Rician likelihood loss for quantitative MRI using self-supervised deep learning}},
  booktitle = {NMR},
  year = {2025},
  doi = {10.1002/nbm.70136},
}
Figure