{"id":"77fcbc38-32a3-4655-88ff-28e2b89fda71","arxiv_id":"2411.15213","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A CDF-matching harmonization method with smooth dual-scaling and tail shrinking matches MRI intensities to a template while preserving local variability.","lead":"This paper introduces an image harmonization method that adjusts MRI intensity distributions to a reference template using constrained curve fitting rather than exact histogram matching. The method is meant to reduce scanner-to-scanner variability while preserving local image features that matter for downstream machine learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The feature-preservation claim is untested at the exact point where the method can demonstrably reduce contrast: the tail-shrinking step compresses high-intensity tails, and the only experimental comparison is against percentile stretch, not exact histogram matching.","rationale":"The reader's weakest assumption is that local CDF deviations encode clinically important features and that suppressing them harms downstream analysis. My concern is a more concrete instance of that assumption: the method's tail-shrinking step provably compresses intensity differences in the upper tail, where many brain pathologies appear, and the paper provides no experiment isolating whether this compression preserves or destroys diagnostic information. The reader's conditional verdict remains appropriate. The mathematics appears coherent, the implementation is linked, and the method may be a reasonable preprocessing tool, so there is no basis for rejection. However, the central advertised benefit is not yet demonstrated: the only downstream experiment compares against percentile stretch, a weak baseline, and does not test the elasticity/tail-shrinking mechanism against the natural alternative of exact histogram matching. A direct feature-preservation experiment, as proposed above, would settle whether the method's distinctive design actually delivers its headline promise.","tokens_in":6591,"tokens_out":4867,"duration_ms":57615,"concrete_test":"On a public brain-tumor MRI dataset (e.g., BraTS), harmonize each exam with: (i) exact histogram matching to the same template, (ii) the proposed restricted CDF matching with the Section 3 parameters, and (iii) the proposed method with tail shrinking disabled (vclipT set beyond vmax). For each setting, compute tumor-to-background contrast-to-noise ratio and train or evaluate the same segmentation architecture to obtain Dice on enhancing tumor and FLAIR hyperintensity. If setting (ii) does not outperform (i), or if Dice/CNR degrades as tail compression increases, the feature-preservation premise and the claimed advantage over exact histogram matching are empirically unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated differentiator over prior histogram matching is that the restricted CDF fit preserves local variability and clinically important features (Abstract; Section 2; Figure 3d). The load-bearing premise is that allowing small global CDF deviations and limiting the transform to smooth dual-scaling protects those features. But the transformation is still a global, monotonic intensity mapping, and the tail-shrinking step actively compresses the upper tail. In the reported configuration (Section 3: vT=3300, vclipT=4095, 12-bit), any intensity above vT is remapped by the erf function in Section 2.2, whose derivative at vT is 2*(vclipT-vT)/(vmax-vT). In the Figure 2b example this is roughly 0.75 and decays rapidly toward zero, so hyperintense structures such as T1ce-enhancing tumor or FLAIR hyperintensities, which typically occupy the upper tail, lose contrast. Thus the method can suppress the very features it claims to preserve. No experiment measures feature preservation or compares against exact histogram matching; Table 1 only compares against percentile stretch, a baseline the paper itself criticizes as unstable. If exact histogram matching achieves equal or better tumor segmentation Dice while the proposed method's tail compression reduces tumor contrast, the central claim fails and the positive Table 1 result could be explained by more robust overall normalization rather than by the proposed elasticity constraints.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6916,"tokens_out":5274,"duration_ms":59841,"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":[{"comment":"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.","section":"Section 2.2 and Figure 3d"},{"comment":"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":"Table 1"},{"comment":"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":"Section 2.4"},{"comment":"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.","section":"Section 3"}],"minor_comments":[{"comment":"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":"Section 2.2 and Eq. (2)"},{"comment":"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.","section":"Section 2.3"},{"comment":"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.","section":"Table 1"},{"comment":"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.","section":"Throughout"},{"comment":"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.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is close to a technical report and its experimental section is thin for a full journal article. The central methodological idea is coherent, but the evidence for the main differentiator, feature preservation, is currently missing. I recommend major revision with targeted experiments: an ablation separating the elastic fit from tail shrinking, a comparison with exact histogram matching, and a quantitative evaluation of local variability or pathology segmentation. The unpublished companion work [13] is used as the main benchmark; the relationship and independence of that evaluation should be clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Roman has a real method here, and the write-up is refreshingly concrete. The restricted CDF matching is a genuine variant of histogram matching: smooth dual-scaling with erf blending, tail shrinking, and a template built from averaged z-scored CDFs. The math is simple enough to implement from the text, and the code is public. That is worth something.\n\nThe reported numbers are plausible but thin. One baseline, 102 exams, no error bars, no significance test. A 0.034 EaD improvement over percentile stretch could easily be noise. More importantly, the paper's distinctive claim—that the method preserves local variability and important features—is never measured. The stress test lands: tail shrinking remaps intensities above vT through an erf whose derivative at vT is about 0.75 and falls toward zero. So hyperintense structures, exactly the T1ce-enhancing tumors and FLAIR hyperintensities the paper says it protects, get compressed. The method might still work well because it normalizes the bulk of the distribution robustly, but the paper does not separate those effects. Adding a comparison against exact histogram matching would settle whether the elasticity constraints earn their keep.\n\nThe experiment compares against percentile stretch, which the paper itself dismisses as unstable. That is a weak baseline for a method that claims superiority over other histogram matching algorithms. I would want to see the same tool trained with exact histogram matching and with Z-score normalization, and ideally a direct feature-preservation test, such as tumor segmentation Dice on harmonized images with and without tail shrinking.\n\nThe writing is honest about the method's parameter tuning, and the control-point approach is intuitive for data engineers. The conclusion overreaches a bit ('always worked exceptionally well'), but that does not undermine the technical content.\n\nBottom line: this is a legitimate contribution to the simple, non-ML harmonization toolbox, and the code availability makes it easy to test independently. For a journal or a serious venue, it needs a stronger experimental section. But it deserves peer review, not a desk reject.","headline":"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.","tokens_in":7403,"tokens_out":3198,"would_cite":true,"duration_ms":37406,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CDF-matching harmonization lifts federated brain MRI Dice to 0.837","keywords":["image harmonization","CDF matching","histogram matching","MRI","federated learning","intensity normalization","brain tumor","curve fitting"],"falsifier":"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.","tokens_in":6416,"feed_emoji":"🧠","tokens_out":5632,"duration_ms":53236,"temperature":0.7,"pith_summary":"The paper proposes an image harmonization method that fits each image's cumulative distribution function (CDF) to a template by curve fitting, using two smooth scaling factors and a uniform shift, followed by error-function-based tail shrinking. The aim is to reduce scanner- and protocol-driven intensity variability across sites while leaving small local distribution deviations intact, because those deviations may encode tumors or other diagnostically relevant features. The author demonstrates the method on structural and diffusion MRI, including a federated brain screening tool, where CDF matching raised exam-averaged Dice from 0.803 to 0.837 and sensitivity from 0.94 to 0.98 versus percentile stretching. The paper argues that the method is generic, intuitive to control through a few percentile-anchored points, and usable out of the box for healthy and pathological cases.","feed_headline":"CDF-matching harmonization lifts federated brain MRI Dice to 0.837","feed_subtitle":"Same federated model scores 0.837 exam Dice with CDF matching vs 0.803 with percentile stretch.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the federated brain MRI screening tool and the experimental comparison in Table 1; it is the downstream task whose Dice, sensitivity, and precision improve.","marker":"[13]"},{"why":"The classical histogram-matching and MRI scale-standardization method that this curve-fitting CDF matching extends and contrasts with.","marker":"[3]"},{"why":"An earlier intensity-warping harmonization method cited as an example of sub-optimal histogram matching that can ruin image features.","marker":"[5]"},{"why":"ComBat is the statistical harmonization baseline mentioned in the introduction, representing the batch-effect category this method competes with in practice.","marker":"[1]"},{"why":"Motivates harmonizing multiple MRI channels, including diffusion channels, and aligning their relative intensity ranges, which the method explicitly targets.","marker":"[12]"}],"fun_headline_variants":["Robust restricted CDF matching boosts federated MRI harmonization","Elastic CDF fitting preserves local variability in image harmonization","Harmonize images without losing local features via restricted CDF matching","Restricted CDF matching lifts federated brain MRI Dice to 0.837"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Robust restricted CDF matching boosts federated MRI harmonization","Elastic CDF fitting preserves local variability in image harmonization","Harmonize images without losing local features via restricted CDF matching","Restricted CDF matching lifts federated brain MRI Dice to 0.837"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000545,"raw_usage":{"total_tokens":2590,"prompt_tokens":913,"completion_tokens":1677,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":1601}},"tokens_in":529,"tokens_out":1677,"duration_ms":13089,"temperature":1.0,"reasoning_tokens":1601,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:38:33.688304+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Method used: EaD G-Dice Sens","cited_arxiv_id":null,"evidence_quote":"Supplies the federated brain MRI screening tool and the experimental comparison in Table 1; it is the downstream task whose Dice, sensitivity, and precision improve."},{"cited_title":"unexpected","cited_arxiv_id":null,"evidence_quote":"The classical histogram-matching and MRI scale-standardization method that this curve-fitting CDF matching extends and contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An earlier intensity-warping harmonization method cited as an example of sub-optimal histogram matching that can ruin image features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ComBat is the statistical harmonization baseline mentioned in the introduction, representing the batch-effect category this method competes with in practice."},{"cited_title":"Dost ´al, T","cited_arxiv_id":null,"evidence_quote":"Motivates harmonizing multiple MRI channels, including diffusion channels, and aligning their relative intensity ranges, which the method explicitly targets."}],"review_version":1}