skip to main content

Title: Cramér–Rao bound‐informed training of neural networks for quantitative MRI

To improve the performance of neural networks for parameter estimation in quantitative MRI, in particular when the noise propagation varies throughout the space of biophysical parameters.

Theory and Methods

A theoretically well‐founded loss function is proposed that normalizes the squared error of each estimate with respective Cramér–Rao bound (CRB)—a theoretical lower bound for the variance of an unbiased estimator. This avoids a dominance of hard‐to‐estimate parameters and areas in parameter space, which are often of little interest. The normalization with corresponding CRB balances the large errors of fundamentally more noisy estimates and the small errors of fundamentally less noisy estimates, allowing the network to better learn to estimate the latter. Further, proposed loss function provides an absolute evaluation metric for performance: A network has an average loss of 1 if it is a maximally efficient unbiased estimator, which can be considered the ideal performance. The performance gain with proposed loss function is demonstrated at the example of an eight‐parameter magnetization transfer model that is fitted to phantom and in vivo data.


Networks trained with proposed loss function perform close to optimal, that is, their loss converges to approximately 1, and their performance is superior to networks trained with the standard mean‐squared error (MSE). The proposed loss function reduces the bias of the estimates compared to the MSE loss, and improves the match of the noise variance to the CRB. This performance gain translates to in vivo maps that align better with the literature.


Normalizing the squared error with the CRB during the training of neural networks improves their performance in estimating biophysical parameters.

more » « less
Award ID(s):
2009752 1922658
Author(s) / Creator(s):
 ;  ;  ;  ;  ;  ;  
Publisher / Repository:
Wiley Blackwell (John Wiley & Sons)
Date Published:
Journal Name:
Magnetic Resonance in Medicine
Page Range / eLocation ID:
p. 436-448
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Purpose

    To develop a scan‐specific model that estimates and corrects k‐space errors made when reconstructing accelerated MRI data.


    Scan‐specific artifact reduction in k‐space (SPARK) trains a convolutional‐neural‐network to estimate and correct k‐space errors made by an input reconstruction technique by back‐propagating from the mean‐squared‐error loss between an auto‐calibration signal (ACS) and the input technique’s reconstructed ACS. First, SPARK is applied to generalized autocalibrating partially parallel acquisitions (GRAPPA) and demonstrates improved robustness over other scan‐specific models, such as robust artificial‐neural‐networks for k‐space interpolation (RAKI) and residual‐RAKI. Subsequent experiments demonstrate that SPARK synergizes with residual‐RAKI to improve reconstruction performance. SPARK also improves reconstruction quality when applied to advanced acquisition and reconstruction techniques like 2D virtual coil (VC‐) GRAPPA, 2D LORAKS, 3D GRAPPA without an integrated ACS region, and 2D/3D wave‐encoded imaging.


    SPARK yields SSIM improvement and 1.5 – 2× root mean squared error (RMSE) reduction when applied to GRAPPA and improves robustness to ACS size for various acceleration rates in comparison to other scan‐specific techniques. When applied to advanced reconstruction techniques such as residual‐RAKI, 2D VC‐GRAPPA and LORAKS, SPARK achieves up to 20% RMSE improvement. SPARK with 3D GRAPPA also improves RMSE performance by ~2×, SSIM performance, and perceived image quality without a fully sampled ACS region. Finally, SPARK synergizes with non‐Cartesian, 2D and 3D wave‐encoding imaging by reducing RMSE between 20% and 25% and providing qualitative improvements.


    SPARK synergizes with physics‐based acquisition and reconstruction techniques to improve accelerated MRI by training scan‐specific models to estimate and correct reconstruction errors in k‐space.

    more » « less
  2. Abstract

    Satellite precipitation products, as all quantitative estimates, come with some inherent degree of uncertainty. To associate a quantitative value of the uncertainty to each individual estimate, error modeling is necessary. Most of the error models proposed so far compute the uncertainty as a function of precipitation intensity only, and only at one specific spatiotemporal scale. We propose a spectral error model that accounts for the neighboring space–time dynamics of precipitation into the uncertainty quantification. Systematic distortions of the precipitation signal and random errors are characterized distinctively in every frequency–wavenumber band in the Fourier domain, to accurately characterize error across scales. The systematic distortions are represented as a deterministic space–time linear filtering term. The random errors are represented as a nonstationary additive noise. The spectral error model is applied to the IMERG multisatellite precipitation product, and its parameters are estimated empirically through a system identification approach using the GV-MRMS gauge–radar measurements as reference (“truth”) over the eastern United States. The filtering term is found to be essentially low-pass (attenuating the fine-scale variability). While traditional error models attribute most of the error variance to random errors, it is found here that the systematic filtering term explains 48% of the error variance at the native resolution of IMERG. This fact confirms that, at high resolution, filtering effects in satellite precipitation products cannot be ignored, and that the error cannot be represented as a purely random additive or multiplicative term. An important consequence is that precipitation estimates derived from different sources shall not be expected to automatically have statistically independent errors.

    Significance Statement

    Satellite precipitation products are nowadays widely used for climate and environmental research, water management, risk analysis, and decision support at the local, regional, and global scales. For all these applications, knowledge about the accuracy of the products is critical for their usability. However, products are not systematically provided with a quantitative measure of the uncertainty associated with each individual estimate. Various parametric error models have been proposed for uncertainty quantification, mostly assuming that the uncertainty is only a function of the precipitation intensity at the pixel and time of interest. By projecting satellite precipitation fields and their retrieval errors into the Fourier frequency–wavenumber domain, we show that we can explicitly take into account the neighboring space–time multiscale dynamics of precipitation and compute a scale-dependent uncertainty.

    more » « less
  3. Abstract

    Multispecies occupancy models estimate dependence among multiple species of interest from patterns of co‐occurrence, but problems associated with separation and boundary estimates can lead to unreasonably large estimates of parameters and associated standard errors when species are rarely observed at the same site or when data are sparse. In this paper, we overcome these issues by implementing a penalized likelihood, which introduces a small bias in parameter estimates in exchange for a potentially large reduction in variance. We compare parameter estimates obtained from both penalized and unpenalized multispecies occupancy models fit to simulated data that exhibit various degrees of separation and to a real‐word data set of bird surveys with little apparent overlap between potentially interacting species. Our simulation results demonstrate that penalized multispecies occupancy models did not exhibit boundary estimates and produced lower bias, lower mean squared error, and improved inference relative to unpenalized models. When applied to real‐world data, our penalized multispecies occupancy model constrained boundary estimates and allowed for meaningful inference related to the interactions of two species of conservation concern. To facilitate the use of our penalized multispecies occupancy model, the techniques demonstrated in this paper have been integrated into theunmarkedpackage in R programing language.

    more » « less
  4. Summary

    We focus on selecting optimal bandwidths for non-parametric estimation of the two-point correlation function of a point pattern. We obtain these optimal bandwidths by using a bootstrap approach to select a bandwidth that minimizes the integrated squared error. The variance term is estimated by using a non-parametric spatial bootstrap, whereas the bias term is estimated with a plug-in approach using a pilot estimator of the two-point correlation function based on a parametric model. The choice of parametric model for the pilot estimator is very flexible. Depending on applications, parametric statistical point models, physical models or functional models can be used. We also explore the use of the procedure for selecting adaptive optimal bandwidths. We investigate the performance of the bandwidth selection procedure by using a simulation study. In our data example, we apply our method to a Sloan Digital Sky Survey galaxy cluster catalogue by using a pilot estimator based on the power law functional model in cosmology. The resulting non-parametric two-point correlation function estimate is then used to estimate a cosmological mass bias parameter that describes the relationship between the galaxy mass distribution and the underlying matter distribution.

    more » « less
  5. Wren, Jonathan (Ed.)
    Abstract Motivation

    Mathematical models in systems biology help generate hypotheses, guide experimental design, and infer the dynamics of gene regulatory networks. These models are characterized by phenomenological or mechanistic parameters, which are typically hard to measure. Therefore, efficient parameter estimation is central to model development. Global optimization techniques, such as evolutionary algorithms (EAs), are applied to estimate model parameters by inverse modeling, i.e. calibrating models by minimizing a function that evaluates a measure of the error between model predictions and experimental data. EAs estimate model parameters “fittest individuals” by generating a large population of individuals using strategies like recombination and mutation over multiple “generations.” Typically, only a few individuals from each generation are used to create new individuals in the next generation. Improved Evolutionary Strategy by Stochastic Ranking (ISRES), proposed by Runnarson and Yao, is one such EA that is widely used in systems biology to estimate parameters. ISRES uses information at most from a pair of individuals in any generation to create a new population to minimize the error. In this article, we propose an efficient evolutionary strategy, ISRES+, which builds on ISRES by combining information from all individuals across the population and across all generations to develop a better understanding of the fitness landscape.


    ISRES+ uses the additional information generated by the algorithm during evolution to approximate the local neighborhood around the best-fit individual using linear least squares fits in one and two dimensions, enabling efficient parameter estimation. ISRES+ outperforms ISRES and results in fitter individuals with a tighter distribution over multiple runs, such that a typical run of ISRES+ estimates parameters with a higher goodness-of-fit compared with ISRES.

    Availability and implementation

    Algorithm and implementation: Github—

    more » « less