Improved Myelin Water Fraction Estimation and Uncertainty Mapping in Multicomponent T2 Relaxometry via NNLS Bootstrap Aggregation

arXiv:2611.14395 · physics.med-ph · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Improved Myelin Water Fraction Estimation and Uncertainty Mapping in Multicomponent T2 Relaxometry via NNLS Bootstrap Aggregation".

Title and authors: Mira: So what’s the core idea of this paper? They are proposing a method to make that MWF estimate more robust by combining residual resampling with repeated NNLS reconstruction.

Kai: That’s right. Instead of just giving you one point estimate, they’re generating variability maps using bootstrap standard deviations from the residual resampling process.

Lev: So, what does this mean for running this on real hardware? It sounds like it adds a lot of overhead because you have to fit the signal repeatedly for every bootstrap replicate.

Kai: That’s a fair point, Lev. The paper notes that the main practical cost is indeed the repeated spectral fitting, but they found that it works well in practice.

Mira: They compared this NNLS Bootstrap Aggregation method against conventional NNLS and other strategies like RNNLS and model-informed machine learning methods, testing everything under different noise conditions.

Lev: And what’s the result of that comparison? Did it actually beat those other methods on the actual estimation accuracy?

Kai: Yes. The paper shows that NNLS Bootstrap Aggregation achieved the lowest mean absolute error and root mean squared error in every simulated noise regime, which is a strong finding.

Mira: That’s good to see because conventional NNLS tends to have higher dispersion in its errors, so this method seems to reduce those errors significantly.

The paper's summary: Kai: So the summary is that they are using this NNLS Bootstrap Aggregation approach to estimate both the MWF and reconstruct the full T2 spectrum under varying noise levels in simulations.

Mira: They specifically highlight that this method improves not just the MWF estimation itself, but also provides voxelwise bootstrap-variability maps for several spectrum-derived metrics, including things like TMW2.

Lev: So, it’s not just about getting a better number for MWF; they are giving you a map of how much that estimate might change depending on the noise in different spots.

Kai: Precisely. They incorporated this technique into their multicomponent T2 reconstruction toolbox, making it available online for others to use with their own data.

Mira: That’s a practical step, moving it from just a theoretical result to something you can actually put into your workflow for analysis.

Lev: From an engineering standpoint, being able to get those variability maps is useful because it tells you exactly where the estimation is most uncertain.

The paper's improvements: Kai: The authors suggest two main improvements over the prior methods. First, they integrate residual bootstrap aggregation into existing NNLS reconstruction pipelines to provide voxelwise variability maps from a single acquisition.

Mira: So they are solving that issue where conventional NNLS only gives you a point estimate without quantifying the voxelwise estimation variability directly.

Lev: That’s what I was thinking. It addresses the fact that you can't just get one number for an entire brain scan; you need to know the local confidence in every voxel.

Kai: They also improved regional scan–rescan repeatability in vivo, showing the method achieved the lowest mean standard deviation of difference and highest mean correlation across two independent cohorts.

Mira: That’s a big deal for clinical applications. It means when you scan someone twice, this method is more consistent across those sessions than what was previously achieved with Bayes-Reg-D minus one or MIML.

Lev: The in vivo results are interesting because they show superior scan–rescan agreement, which suggests the method is robust enough for real patient data.

Conclusion: Kai: So to wrap up the paper "Improved Myelin Water Fraction Estimation and Uncertainty Mapping in Multicomponent T2 Relaxometry via NNLS Bootstrap Aggregation", they confirm that this approach is useful for improving MWF estimation while providing voxelwise variability maps from the same acquisition.

Mira: They show that it achieved the lowest mean absolute error and root mean squared error across all simulated noise regimes compared to conventional NNLS, and it also had lower dispersion in its errors.

Lev: And they found that this method yielded the most favorable results for regional scan–rescan repeatability in two independent cohorts, with Cohort one showing the best performance <ref:2611.14395#pg1>.

Kai: The paper concludes that while repeated spectral fitting is the main practical cost, the results support residual-bootstrap aggregation as a useful approach to MWF estimation.

Mira: And they also found that diminishing improvements were seen as the number of replicates increased, meaning you don't need an infinitely large ensemble to get very good estimates.

Lev: So it sounds like the message is that this technique gives you better estimates and better uncertainty maps for myelin water fraction without needing the absolute most accurate reconstruction of the full spectrum.

Erick J Canales-Rodríguez, Thomas Yu, Juan Manuel Górriz, Santiago Galdo-Álvarez, Tobias Kober, Tom Hilbert, Gian Franco Piredda, Miguel Ángel Rivas-Fernández

Department of Signal Theory, Networking and Communications, University of Granada · Research Centre for Information and Communication Technologies (CITIC), University of Granada Research Centre for Information and Communication Technologies (CITIC) · Andalusian Research Institute in Data Science and Computational Intelligence (DaSCI), Andalusian Research Institute in Data Science and Computational Intelligence (DaSCI) · Swiss Innovation Hub, Siemens Healthineers International AG · Department of Clinical Psychology and Psychobiology, University of Santiago de Compostela · Cognitive Neuroscience Research and Psychogerontology Group (NeuCogA-Aging), Institute of Psychology (IPsiUS), University of Santiago de Compostela · Health Research Institute of Santiago de Compostela, Health Research Institute of Santiago de Compostela · Research & Clinical Translation, Magnetic Resonance, Siemens Healthineers AG · Department of Psychology, Sociology and Philosophy, University of León

physics.med-ph

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: 22 pages, 5 figures, 3 tables

Code: https://github.com/ejcanalesr/multicomponent-T2-toolbox

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Estimating the myelin water fraction (MWF) from multicomponent T2 relaxometry requires solving an ill-conditioned inverse problem As a result…

Key concepts

Multicomponent T2 Relaxometry
This MRI technique separates the tissue signal into components based on different transverse relaxation properties (T2 values). By analyzing these components, researchers aim to determine the myelin water fraction (MWF), which is a key measure of myelin content in brain tissue.
Ill-conditioned Inverse Problem
Estimating MWF from T2 data is mathematically difficult because components with similar T2 values produce very similar signal decays. This makes the inverse problem unstable, meaning small measurement noises can cause large errors in the calculated MWF.
NNLS Bootstrap Barycenter (NNLS-BB)
This proposed method uses residual resampling to create slightly perturbed signals and then fits each with NNLS. Averaging these results across many bootstrap replicates provides a more robust estimate, significantly lowering the mean absolute error and root mean squared error compared to conventional methods.

Terminology

Summary

Estimating the myelin water fraction (MWF) from multicomponent T2 relaxometry requires solving an ill-conditioned inverse problem As a result, MWF estimates can be highly sensitive to measurement noise The proposed method is the nonnegative least squares Bootstrap Barycenter (NNLS-BB), which combines residual resampling with repeated NNLS reconstruction This approach requires neither an explicit spectral regularization penalty nor a training dataset

Problem Context

Multicomponent T2 relaxometry probes brain tissue microstructure by separating MRI signal contributions associated with water pools exhibiting different transverse relaxation properties The relative contribution of this short-T2 component defines the myelin water fraction (MWF), a widely studied MRI marker associated with myelin content A conventional approach models the measured signal as a superposition of exponential relaxation components characterized by different T2 values This inverse problem is intrinsically ill-conditioned because components with nearby T2 values generate very similar signal decays over the finite set of measured echo times The instability is particularly problematic for myelin water because the short-T2 component decays rapidly and represents only a small fraction of the total signal

Methodology

The NNLS-BB method is implemented by generating perturbed signals through residual resampling and fitting each signal with NNLS For each measured signal y, the authors first compute xbnnls using Eq. (2) The raw residuals are then mean-centered and subjected to a standard leverage adjustment before bootstrap resampling For each bootstrap replicate b = 1,. p. 4.., B, a perturbed signal y∗(b) is constructed by adding randomly sampled entries from the adjusted residuals to the original NNLS fit The resulting spectra and derived parameter values are averaged after MWF-based trimming using retained replicates

Performance Metrics

NNLS-BB achieved the lowest MWF mean absolute error and root mean squared error in every simulated noise regime Mean absolute error decreased by 36–41% relative to conventional NNLS, alongside lower error dispersion Compared with Bayes-Reg-D−1 and MIML, NNLS-BB achieved comparable dispersion but substantially lower bias In vivo, NNLS-BB yielded the lowest regional scan–rescan dispersion and highest mean correlation in both cohorts

In Vivo Validation

Scan–rescan repeatability was assessed across 20 white-matter regions in two independent cohorts comprising 13 healthy participants NNLS-BB achieved the most favorable values across all reported measures in Cohort 1 Across both cohorts, NNLS-BB provided the lowest mean SDdiff, the highest mean correlation, and regression slopes closest to unity

Conclusion

These results support residual-bootstrap aggregation as a practical approach to improving MWF estimation while providing voxelwise variability maps from the same acquisition The main practical cost of NNLS-BB is the repeated spectral fitting Overall, the results support NNLS residual-bootstrap aggregation as a useful approach to MWF estimation The ensemble-size analysis revealed diminishing improvements as the number of replicates increased, with B = 200 producing estimates close to those obtained with substantially larger ensembles The study concludes that improved MWF estimation did not require the most accurate reconstruction of the full spectrum

--- Page 26 ---

REFERENCES

Eva Alonso-Ortiz, Ives R. Levesque, and G. Bruce Pike. MRI-based myelin water imaging: A technical review Trevor Andrews, Jack L. p. 1 Lancaster, Stephen J. Dodd, Carmen Contreras-Sesvold, and Peter T. Fox. Testing the three-pool white matter model adapted for use with T2 relaxometry B. p. 6 B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain M Bertero, P Boccacci, and Edward Roy Pike. p. 8 On the recovery and resolution of exponential relaxation rates from experimental data: a singular-value analysis of the Laplace transform inversion in the presence of noise A D Parsekian and D Grombacher. Uncertainty estimates for surface nuclear magnetic resonance water content and relaxation time profiles from bootstrap statistics J V Manjon, P Coupe, L Concha, A Buades, D L Collins, and M Robles. p. 0 Diffusion weighted image denoising using overcomplete local PCA Sandra M Meyers, Shannon H Kolind, and Alex L MacKay. p. 8 Simultaneous measurement of total water content and myelin water fraction in brain at 3 T using a T2 relaxation based method Karl E Taylor. p. 5 Summarizing multiple aspects of model performance in a single diagram Matthias Weigel. p. 2 Extended phase graphs: Dephasing, RF pulses, and echoes—pure and simple Kenneth P Whittall and Alexander L MacKay. p. 2 Quantitative interpretation of NMR relaxation data Alex Mackay, Kenneth Whittall, Julian Adler, David Li, Donald Paty, and Douglas Graeb. In vivo visualization of myelin water in brain by magnetic resonance

--- Page 27 ---

Figure 1: Effect of the number of bootstrap replicates on the mean absolute error (MAE) of the estimated myelin water fraction (MWF). The left panel shows the MWF MAE for each SNR range as a function of the number of bootstrap replicates, from 0 to 1000. p. 9 The right panel shows the ratio between the MWF MAE obtained with each number of replicates and the value obtained with 1000, providing a normalized measure of convergence toward the largest ensemble size evaluated NNLS, NNLS-BB reduced cRMSE by 36–43% across SNR regimes, indicating less dispersion in MWF estimation errors Relative to its strongest competitors, Bayes-Reg-D−1 and MIML, its cRMSE was generally comparable, but the average magnitude of its MBE across the four SNR regimes was approximately 82% and 86% lower, respectively

--- Page 28 ---

Figure 2: Voxelwise absolute myelin water fraction (MWF) error distributions across SNR ranges. Across all SNR regimes, NNLS-BB shifted the error distribution toward lower values compared to the classical NNLS-based methods and MIML This was reflected in smaller median and mean errors, in agreement with Table 2, and in more compact interquartile ranges

--- Page 29 ---

Figure 3: Illustration of the NNLS-BB estimator for two representative voxels. The upper row corresponds to a voxel from the lowest SNR regime (SNR = 50−100), whereas the lower row corresponds to a voxel from the highest SNR regime (SNR = 400−1000) The upper row shows that the bootstrap mean does not perfectly recover the position of the short-T2 lobe, which appears shifted toward shorter relaxation times This illustrates an important feature of the problem: shifts in the location of the myelin-related lobe do not necessarily imply a large error in the estimated MWF

--- Page 30 ---

Figure 4: Scan–rescan agreement of NNLS-BB at the individual subject level. The axial MWF maps show similar spatial patterns across sessions, with local differences in estimated amplitudes The Bland–Altman plots summarize the corresponding regional measurements for all subjects in each cohort Cohort 2 showed less scan–rescan dispersion than Cohort 1, consistent with the lower SDdiff values in Table 3

--- Page 30 ---

Figure 5: Spectrum-derived parameter maps from NNLS-BB, bootstrap variability, and reconstructed white-matter spectra. The left panel shows axial maps of MWF, IEWF, quasi-free water fraction, TMW2, TIE2, and TWC The right panel shows reconstructed spectra from voxels within a white-matter mask

--- Page 30 ---

Table 3: Scan–rescan comparison of MWF estimates in the two in vivo cohorts. Metrics were calculated for each subject across the white-matter tract ROIs and then summarized across subjects NNLS-BB achieved the most favorable values across all reported measures in Cohort 1

--- Page 30 ---

Figure 4: Scan–rescan MWF estimates obtained with NNLS-BB. The left and right panels correspond to Cohorts 1 and 2 respectively All five subjects from Cohort 1 are displayed; five of the eight subjects from Cohort 2 are shown for a balanced visual layout

--- Page 30 ---

Figure 5: Spectrum-derived parameter maps from NNLS-BB, bootstrap variability, and reconstructed white-matter spectra. The left panel shows axial maps of MWF, IEWF, quasi-free water fraction, TMW2, TIE2, and TWC

--- Page 30 ---

Figure 4: Scan–rescan MWF estimates obtained with NNLS-BB. The left and right panels correspond to Cohorts 1 and 2 All five subjects from Cohort 1 are displayed; five of the eight subjects from Cohort 2 are shown for a balanced visual layout

--- Page 30 ---

Figure 4: Scan–rescan MWF estimates obtained with NNLS-BB.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Residual bootstrap aggregation can be integrated into existing NNLS reconstruction pipelines to provide voxelwise variability maps from a single acquisition, addressing the limitation that conventional NNLS methods provide point estimates without directly quantifying voxelwise estimation variability.

  2. The improved AI system can estimate myelin water fraction (MWF) with lower mean absolute error and root mean squared error across all simulated noise regimes compared to conventional NNLS, as shown by the finding that NNLS-BB achieved the lowest MWF mean absolute error and root mean squared error in every simulated noise regime.

  3. The system can provide voxelwise variability maps for spectrum-derived metrics, such as MWF and TMW2, using bootstrap standard deviations derived from residual resampling, which characterize the parameter-dependent variability captured by the bootstrap ensemble.

  4. The AI system can yield favorable regional scan–rescan repeatability in in vivo settings by achieving the lowest mean SDdiff and highest mean correlation across two independent cohorts, supporting its applicability beyond a single dataset.

Related papers