Rician Likelihood Loss for Quantitative MRI With Self-Supervised Deep Learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Biophysics

Background:

• Quantitative MRI enables robust parameter estimation without training labels using self-supervised deep learning.
• Existing methods using mean squared error (MSE) loss show significant bias at low signal-to-noise ratios (SNR).
• MSE loss is incompatible with MR magnitude signals, potentially causing estimation bias.

Purpose of the Study:

• Introduce a novel Rician likelihood loss (NLR) for self-supervised learning in quantitative MRI.
• Address the bias issue in parameter estimation caused by low SNR.
• Improve the accuracy and robustness of quantitative MRI.

Main Methods:

• Developed a stable numerical approximation for the negative log Rician (NLR) likelihood loss.
• Compared NLR loss against traditional MSE loss using the intravoxel incoherent motion (IVIM) model.
• Evaluated parameter estimation performance using simulated and real MRI data across various SNRs.

Main Results:

• NLR loss significantly reduced bias in IVIM diffusion coefficient estimates at low SNR.
• At low SNR, NLR loss improved accuracy at the cost of precision.
• Performance converged for NLR and MSE losses at higher SNRs, yielding higher accuracy, precision, and lower total error.

Conclusions:

• The NLR loss enhances accuracy in quantitative MRI parameter estimation, especially in noisy conditions.
• This method offers broad applicability for improving MRI analysis from low-SNR data.
• The NLR loss provides a more robust alternative to MSE loss for self-supervised MRI.