Input layer regularization for magnetic resonance relaxometry biexponential parameter estimation

Area of Science:

• Physics
• Biophysics
• Computational Science

Background:

• Estimating parameters for biexponential decay signals is crucial in magnetic resonance relaxometry (MRR) and physical sciences.
• This parameter estimation is an intrinsically ill-posed problem, making estimates sensitive to noise and parameter values.
• Regularization and neural networks are existing methods for addressing ill-posed problems in parameter estimation.

Purpose of the Study:

• To introduce a novel neural network regularization technique called input layer regularization (ILR) for biexponential decay models.
• To evaluate the effectiveness of ILR in improving the accuracy of time constant estimates in biexponential decay signals.

Main Methods:

• Developed a novel neural network regularization method termed input layer regularization (ILR).
• ILR augments neural network inputs with signals derived from regularized nonlinear least-squares estimates of decay time constants.
• Compared ILR performance against existing methods for parameter estimation in biexponential decay signals.

Main Results:

• ILR demonstrated a significant reduction in the error of time constant estimates, ranging from 15% to over 50%.
• Greater improvements were observed for the time constant of the more rapidly decaying exponential component.
• The performance enhancement was dependent on the chosen error metric and the signal-to-noise ratio of the data.

Conclusions:

• Input layer regularization (ILR) offers a powerful and efficient method for improving parameter estimation in biexponential models.
• ILR is compatible with existing regularization techniques and shows broad applicability to various parameter estimation challenges.
• This method enhances the reliability of parameter estimates in magnetic resonance relaxometry and other scientific fields.