REVIEW 4 major objections 5 minor 26 references
Image Harmonization using Robust Restricted CDF Matching
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read CDF-matching harmonization lifts federated brain MRI Dice to 0.837
desk verdict A clear, implementable restricted CDF-matching method with shipped code; the feature-preservation claim is plausible but unsupported by the single-baseline experiment, and tail-shrinking may compress the exact hyperintensities it promises to protect. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the CDF-matching curve fit: for each input image, its CDF $CDF_I$ is fitted to a template $CDF_T$ by optimizing two scaling factors $\sigma_B$, $\sigma_T$ and a shift $\gamma$ in the lookup table $LUT_{ds}(x) = (x - v_M)\cdot\sigma_\kappa(x,\sigma_B,\sigma_T) + \gamma$, where the pointwise scaling $\sigma_\kappa$ blends $\sigma_B$ and $\sigma_T$ with a sigmoidal error function around a pivot intensity $v_M$ to avoid a sharp artifact. Tail shrinking then maps the remaining upper and lower tails with $LUT_T(x) = v_T + r_T \operatorname{erf}(2(x-v_T)/r_S)$ and its mirrored counterpart, so values stay within a target range. The template is built as an average CDF of z-score-normalized training images and anchored to three control points $\pi_B, \pi_M, \pi_T$; dual-scaling guarantees a suitable fit. This machinery does the work of harmonizing while explicitly allowing tiny local deviations from the template to survive, which the paper identifies with preserved features.
What would settle it
Run the same federated screening tool on images harmonized with a deliberately over-constrained template that forces the CDF to match almost exactly, allowing no local deviations; if exam-averaged Dice and sensitivity stay at the 0.837 and 0.98 levels, then the preservation of local deviations is not what drives the improvement, and the method's core premise is unsupported. An ablation that disables tail shrinking while keeping dual-scaling would also reveal how much of the gain comes from range clipping rather than from elastic CDF fitting.
Extended reading notes
Core claim
The paper's central claim is that harmonization should be formulated as a robust, restricted CDF-matching problem rather than as exact histogram equalization or unconstrained machine-learning mapping. Under this formulation, the image CDF is only loosely coupled to a pre-defined template CDF: the fitting procedure permits a smooth dual scaling of the bottom and top halves of the distribution around a pivot, blended by an error function, plus a constant offset, and then shrinks long tails with an erf-based map to fit a target intensity range. This loose coupling is what preserves local deviations, and the same mechanism makes the template easy to define from a few control points, such as target intensities at the median and upper and lower percentiles. The supporting experiment shows that a federated learning brain MRI screening tool trained and tested on CDF-harmonized images outperforms the same tool on percentile-stretched images on all reported metrics.
Load-bearing premise
The claim depends on the assumption that small local differences between an image's intensity distribution and the template are meaningful diagnostic features, such as tumors, rather than noise, so preserving them is what improves downstream performance.
Editorial extensions
If this is right
- A machine-learning model trained on CDF-harmonized images can outperform one trained on percentile-stretched images: the federated brain screening tool reports exam-averaged Dice 0.837 versus 0.803, global Dice 0.884 versus 0.852, sensitivity 0.98 versus 0.94, and precision 0.91 versus 0.90.
- Because the template is defined independently and anchored by just a few control points, the same harmonizer can be reused across channels and sites without retraining, and the relative intensity ranges of different MRI channels can be aligned to avoid channel bias.
- The method can map images into a fixed target intensity range, such as 12-bit, via tail shrinking while keeping most of the distribution elastic, which suits storage requirements and network-friendly dynamic ranges.
- Since the fit is smooth and elastic, local deviations from the template remain visible after harmonization, so features such as pathological regions are not flattened into the template distribution.
Reading between the lines
- If local deviations from the template are indeed preserved, the residual between a harmonized image and the template could itself serve as an anomaly or pathology map; a testable extension is to measure tumor detection directly on such residuals.
- The claimed insensitivity to template parameters suggests a practical auto-calibration rule: set the median control point from training data and let dual-scaling absorb site differences, which could be validated by a cross-site experiment where only the middle control point is tuned.
- Because the machinery operates on CDFs rather than on spatial structure, the same curve-fitting construction could apply to non-image one-dimensional distributions, such as spectra or flow-cytometry histograms, where harmonizing batch effects matters.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an analytical MRI harmonization method that aligns each image's cumulative distribution function (CDF) to a template by solving a constrained curve-fitting problem. The allowed transformations are a smooth dual scaling of the intensity axis around a pivot, a uniform shift, and an erf-based tail shrinking operation. The template CDF is built from averaged Z-score-normalized training CDFs anchored at three control points. The authors report that using this preprocessing improves a federated brain MRI screening tool over percentile stretching (Table 1: exam-averaged Dice 0.837 versus 0.803), and they argue that the loose CDF fit preserves local variability and clinically important features that exact histogram matching would destroy.
Significance. The method is simple, interpretable, and described in enough implementable detail: the equations, control points, and clipping parameters are explicit, and a public GitLab implementation is referenced. If the reported performance gain is robust and the feature-preservation claim were validated, the method would be a useful lightweight alternative to ML-based harmonization, particularly in federated and multi-site settings. The paper's strengths are its transparent parameterization and its avoidance of black-box models. However, the central claim that the method preserves local variability and important image features is, at present, asserted rather than demonstrated, and the quantitative evidence is limited to a single baseline with no uncertainty quantification.
major comments (4)
- [Section 2.2 and Figure 3d] The paper's central differentiator, stated in the Abstract and Section 2, is that the method preserves local variability and clinically important features. This claim is not measured directly. The proposed transformation is a global, monotonic intensity mapping, and the tail-shrinking function in Section 2.2 compresses the upper tail by design: the erf saturates above vT, reducing contrast in hyperintense structures such as T1ce-enhancing tumor or FLAIR hyperintensities. No experiment tests feature preservation, local variability, or pathology detection as a function of the elasticity of the fit; Figure 3d only shows illustrative CDFs. Because the method's advantage over exact histogram matching is argued to rest on this property, the absence of a quantitative test is load-bearing.
- [Table 1] The quantitative evaluation compares the proposed CDF matching only against percentile stretching, on a single in-house screening tool and a single test set of 102 examinations. No confidence intervals, standard errors, or significance tests are reported, and no multiple-run variability is given. There is also no comparison against exact histogram matching, Z-score normalization, ComBat, or other harmonization methods. Consequently, the headline improvement (EaD 0.837 versus 0.803) could be due to the overall normalization, the clipping, or the template choice rather than to the proposed elastic dual-scaling constraints. An ablation that isolates the dual-scaling/shift step from the tail shrinking, and a comparison with an unconstrained CDF match, are needed to support the central claim.
- [Section 2.4] The curve-fitting optimization is not fully specified: the objective function used to fit CDFI to CDFT is not stated, nor are the initialization, the solver, the convergence criterion, or the handling of ties and background masks. These details affect the fitted parameters sigma_B, sigma_T, and gamma, and therefore the actual harmonized intensities. Since the authors argue that the method's robustness is a key advantage, the optimization procedure should be described precisely or the reader should be pointed to the relevant part of the public implementation.
- [Section 3] The paper claims the method works for both healthy and pathological cases, but the 102-examination test set is not characterized by pathology type, and no per-subgroup results are reported. Given that the method's motivation is preserving pathological features, results broken down by healthy versus tumor-bearing examinations, or by lesion type, would directly address the risk that tail compression damages exactly the structures the method is intended to preserve.
minor comments (5)
- [Section 2.2 and Eq. (2)] The definition of the top-tail shrinking function uses the symbol vclip in the expression rT = vclip - vT while the function signature uses vclipT; please unify the notation. Also, the placement of the factor 2 inside the erf argument is ambiguous in the typeset equation and should be clearly written as 2(x - vT)/rS.
- [Section 2.3] The template CDF preparation first applies Z-score normalization, a method the paper criticizes in the introduction. The averaging argument is plausible, but it should be stated explicitly that Z-score normalization is used only to bring different images onto a comparable scale before averaging, and why this does not reintroduce the instability criticized earlier.
- [Table 1] The metrics EaD, G.Dice, sensitivity, and precision are reported without defining the computation of global Dice or the unit of analysis for sensitivity and precision; please specify them in the text or in a caption.
- [Throughout] The title and abstract use the word 'robust,' but no robustness analysis is provided, such as sensitivity to the control points piB, piM, piT, to the clipping values, or to outliers in the intensity distribution; a brief sensitivity study would substantiate the term.
- [Introduction] There is a typo in the introduction where 'alteration' appears as 'alternation'; also, the reference to Nyul et al. should have the author name spelled consistently as Nyul.
Circularity Check
No significant circularity: the only tautological element is the by-construction closeness of fitted CDFs to the template; the downstream benchmark is external.
-
self definitional
[Section 2.4, Image transformation (recipe steps 2–3 and closing note)]
"Fit CDFI to CDFT using curve fitting optimization, where we allow transformation of CDFI in terms of double-scaling and uniform shifting. ... Please note that this process actually solves two tasks simultaneously: i) it harmonizes images because their intensity ranges will closely follow the template CDFT"
The closeness of the harmonized CDFs to the template is the direct output of the optimization, not an empirical discovery. The per-image parameters σB, σT, and γ are solved by explicitly fitting CDFI to CDFT, so the harmonized intensities 'closely follow' the template by construction. Figure 3d therefore illustrates the fitting objective rather than an independent prediction. This is a minor self-definitional statement about the method's behavior. It does not affect the downstream segmentation comparison, which uses an external network and independent ground-truth-derived Dice, sensitivity, and precision metrics.
full rationale
The paper's core contribution is algorithmic: a constrained curve-fitting CDF-matching transform with dual scaling and tail shrinking. The statement that the transformed images 'closely follow the template CDF' is a restatement of the fitting objective, so it is tautological and accounts for the small score of 2. However, the central scientific evidence is external: Table 1 compares the federated Brain MRI Screening Tool with CDF-matching versus percentile stretch on independent metrics (exam-averaged Dice, global Dice, sensitivity, precision), with no shared fitted parameters between the harmonization transform and the evaluated segmentation outcome. The feature-preservation claim is asserted rather than tested, but that is a validity risk, not circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known results. Thus the paper is essentially self-contained against an external benchmark, and only a by-construction description of the fit keeps the score above zero.
Assumptions & free parameters
free parameters (3)
- Control points πB, πM, πT (target percentiles and intensities) =
πB=(0.1, 500), πM=(0.5, 1650), πT=(0.99, 3300) in the demo; user-chosen in general
- Clipping values vclipB and vclipT =
vclipB=1, vclipT=4095 in the demo
- Per-image scaling factors σB, σT and shift γ =
Not reported; obtained per image by curve fitting
assumptions (3)
- domain assumption The CDF of a standardized MRI image is a meaningful representation of image content, and averaging CDFs from reference images produces a valid template.
- domain assumption Local deviations of individual CDFs from the template represent important features such as pathologies and should be preserved.
- standard math A standard numerical optimizer reliably finds σB, σT, and γ for each image.
Cite this review
Pith. "Pith review of Image Harmonization using Robust Restricted CDF Matching." pith.science (2026). https://pith.science/paper/BGBY4GWE
@misc{pith2026241115213,
author = {Pith},
title = {Pith review of: Image Harmonization using Robust Restricted CDF Matching},
year = {2026},
howpublished = {\url{https://pith.science/paper/BGBY4GWE}},
note = {Machine review of arXiv:2411.15213}
}
read the original abstract
Deployment of machine learning algorithms into real-world practice is still a difficult task. One of the challenges lies in the unpredictable variability of input data, which may differ significantly among individual users, institutions, scanners, etc. The input data variability can be decreased by using suitable data preprocessing with robust data harmonization. In this paper, we present a method of image harmonization using Cumulative Distribution Function (CDF) matching based on curve fitting. This approach does not ruin local variability and individual important features. The transformation of image intensities is non-linear but still ``smooth and elastic", as compared to other known histogram matching algorithms. Non-linear transformation allows for a very good match to the template. At the same time, elasticity constraints help to preserve local variability among individual inputs, which may encode important features for subsequent machine-learning processing. The pre-defined template CDF offers a better and more intuitive control for the input data transformation compared to other methods, especially ML-based ones. Even though we demonstrate our method for MRI images, the method is generic enough to apply to other types of imaging data.
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION Medical imaging, particularly magnetic resonance imaging (MRI), has become an indispensable tool in clinical practice and research, providing detailed insights into anatomical struc- tures and physiological processes. However, the variability in MRI data arising from differences in acquisition proto- cols, scanner hardware, or cross-instituti...
-
[2]
METHOD The main idea of our approach is to match the image CDF function to a pre-defined template CDF (target). The main difference from other histogram matching methods is that we solve the CDF-matching as a curve-fitting problem with post- processing to shrink the long tails of the distribution. We are not modifying the histogram to “perfectly” match th...
work page 2000
-
[3]
adjusts the intensity distribution of an image to resemble that of a reference dataset or reference distribution. There have been multiple proposed algorithms, e.g., White Stripe harmonization [4], RA VEL [5], mica [6] or RIDA [7]. Machine learning approaches have gained traction in recent years, with techniques such as generative adversarial networks (GA...
-
[4]
Our implementation for Python is available in GitLab1
Apply tail-shrinking LU TT and LU TB if the concrete ap- plication requires constrained intensity values. Our implementation for Python is available in GitLab1. Please note that this process actually solves two tasks simultaneously: i) it harmonizes images because their intensity ranges will closely follow the template CDFT , and ii) it can increase con- ...
work page 2000
-
[5]
Compute the image CDF I for input image I
-
[6]
A fitted solution will give us optimal values of parameters σB, σT , and γ
Fit CDFI to CDFT using curve fitting optimization, where we allow transformation of CDFI in terms of double- scaling and uniform shifting. A fitted solution will give us optimal values of parameters σB, σT , and γ
-
[7]
1 to transform the image intensities
For obtained parameters σB, σT , and γ, use LU Tds func- tion from Eq. 1 to transform the image intensities
-
[8]
Mengting Liu et al., “Style transfer generative adversarial networks to harmonize multisite mri to a single reference image to avoid overcorrection,” Human Brain Mapping, vol. 44, no. 14, pp. 4875–4892, 2023
work page 2023
Show all 26 references
-
[9]
EXPERIMENTS AND RESULTS We tested this method extensively for MRI images of the brain, both for patients with tumors (256) and healthy patients (56). When we employed this harmonization algorithm for our Brain MRI Screening Tool [13], which works in federated environ- ment, th...
-
[10]
and ESPA [11]. In contrast to other recent methods that focus on model- ing and/or correcting the scanner effect, or which study the harmonization from more theoretical point of view, our aim is clearly different: we want to present a technically sound and robust method, still...
2024 arXiv
-
[11]
CONCLUSION We presented our proposed image harmonization method based on the CDF matching using curve fitting. Curve fitting allows for local image deviations from template CDF, which i) helps preserve important image features for subsequent processing, and ii) allows for lowe...
-
[12]
Dost ´al, T
ACKNOWLEDGMENTS The author would like to thank our colleagues M. Dost ´al, T. Kopˇrivov´a, and M. Ke ˇrkovsk´y for providing MRI data on which this method was demonstrated, and Prof. Michal Kozubek for his support of this development
-
[13]
Method used: EaD G-Dice Sens
with our CDF-matching harmonization method, as compared to the results obtained by exactly the same network trained and tested on input images harmonized using simple percentile stretch. Method used: EaD G-Dice Sens. Prec. CDF-matching 0.837 0.884 0.98 0.91 Percentile stretch ...
-
[14]
Combat harmonization: Em- pirical bayes versus fully bayes approaches,
Maxwell Reynolds et al., “Combat harmonization: Em- pirical bayes versus fully bayes approaches,” NeuroIm- age: Clinical, vol. 39, pp. 103472, 2023
2023
-
[15]
Adjusting batch effects in mi- croarray expression data using empirical bayes methods,
W Evan Johnson et al., “Adjusting batch effects in mi- croarray expression data using empirical bayes methods,” Biostatistics, vol. 8, no. 1, pp. 118–127, 2007
2007
-
[16]
New variants of a method of mri scale standardization,
L´aszl´o G Ny´ul et al., “New variants of a method of mri scale standardization,” IEEE transactions on medical imaging, vol. 19, no. 2, pp. 143–150, 2000
2000
-
[17]
Statistical normalization techniques for magnetic resonance imaging,
Russell T Shinohara et al., “Statistical normalization techniques for magnetic resonance imaging,” NeuroIm- age: Clinical, vol. 6, pp. 9–19, 2014
2014
-
[18]
Removing inter-subject tech- nical variability in magnetic resonance imaging studies,
Jean-Philippe Fortin et al., “Removing inter-subject tech- nical variability in magnetic resonance imaging studies,” NeuroImage, vol. 132, pp. 198–212, 2016
2016
-
[19]
Intensity warping for multisite mri harmonization,
J. Wrobel et al., “Intensity warping for multisite mri harmonization,” NeuroImage, vol. 223, 2020
2020
-
[20]
A robust intensity distribu- tion alignment for harmonization of t1w intensity values,
Donatas Sedereviˇcius et al., “A robust intensity distribu- tion alignment for harmonization of t1w intensity values,” bioRxiv, p. 2022.06.15.496227, 6 2022
2022
-
[21]
Unsupervised mr harmonization by learning disentangled representations using information bottleneck theory,
Lianrui Zuo et al., “Unsupervised mr harmonization by learning disentangled representations using information bottleneck theory,” NeuroImage, vol. 243, pp. 118569, 2021
2021
-
[22]
Mispel: A su- pervised deep learning harmonization method for multi- scanner neuroimaging data,
Mahbaneh Eshaghzadeh Torbati et al., “Mispel: A su- pervised deep learning harmonization method for multi- scanner neuroimaging data,” Medical image analysis , vol. 89, pp. 102926, 2023
2023
-
[23]
Espa: An un- supervised harmonization framework via ˆA enhanced structure preserving augmentation,
Mahbaneh Eshaghzadeh Torbati et al., “Espa: An un- supervised harmonization framework via ˆA enhanced structure preserving augmentation,” in Medical Image Computing and Computer Assisted Intervention – MIC- CAI 2024, Cham, 2024, pp. 184–194, Springer Nature Switzerland
2024
-
[24]
Possibilities of using multi-b-value diffusion magnetic resonance imaging for classification of brain lesions,
Tereza Kopˇrivov´a, Miloˇs Keˇrkovsk´y, Tom´aˇs J˚uza, V´aclav Vyb´ıhal, Tom´aˇs Rohan, Michal Kozubek, and Marek Dost´al, “Possibilities of using multi-b-value diffusion magnetic resonance imaging for classification of brain lesions,” Academic Radiology, vol. 31, no. 1, pp. 2...
2024
-
[25]
Brain mri screening tool with federated learning,
Roman Stoklasa et al., “Brain mri screening tool with federated learning,” in 2024 IEEE International Sym- posium on Biomedical Imaging (ISBI) . IEEE, 2024, pp. 1–5
2024
-
[26]
The MRI data used for the demonstration of this method come from the study, which was performed in line with the principles of the Declaration of Helsinki
COMPLIANCE WITH ETHICAL STANDARDS This is a mathematical study for which no ethical approval was required. The MRI data used for the demonstration of this method come from the study, which was performed in line with the principles of the Declaration of Helsinki. Approval was g...
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.