Bias-reduced neural networks for parameter estimation in quantitative MRI

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medical Imaging
• Quantitative Imaging

Background:

• Quantitative MRI parameter estimation is crucial for medical diagnosis.
• Traditional methods like nonlinear least-squares fitting can be computationally intensive and prone to bias.
• Neural networks (NNs) show promise but often struggle with bias and variance control.

Purpose of the Study:

• To develop NN-based quantitative MRI parameter estimators with minimal bias.
• To achieve estimation variance close to the theoretical Cramér-Rao bound.
• To improve the accuracy and efficiency of MRI parameter mapping.

Main Methods:

• Generalized the mean squared error loss function to control NN bias and variance.
• Incorporated averaging over multiple noise realizations during NN training.
• Evaluated NN performance in simulations and two in vivo neuroimaging applications.

Main Results:

• The proposed NN strategy significantly reduced estimation bias across the parameter space.
• Achieved estimation variance close to the Cramér-Rao bound in simulations.
• Demonstrated good concordance with traditional estimators in vivo, outperforming state-of-the-art NNs.

Conclusions:

• The developed NNs exhibit substantially reduced bias compared to standard MSE-trained NNs.
• Offer superior computational efficiency over traditional MRI parameter estimation techniques.
• Provide comparable or improved accuracy, representing a significant advancement in quantitative MRI.