{"id":"86391e6f-7e2d-4195-ab5d-5fdf5a6f4620","arxiv_id":"2509.02952","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"STAR aligns multi-stain serial whole-slide pathology images through hierarchical rotation-translation correlation, with stain-conditioned preprocessing and quality control.","lead":"STAR is a new open-source tool for rigidly aligning serial whole-slide histopathology images across different stains. It uses fast correlation-based search with stain-aware preprocessing and is claimed to align slides within minutes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative evaluation: Section IV describes datasets but reports no error metrics or baseline comparisons, so the abstract's claim of robust, stable alignment on ANHIR/ACROBAT is unsupported.","rationale":"The reader's verdict REJECT is, in my view, correct. The strongest_claim is exactly the empirical claim that no experiment supports. My focus differs slightly from the reader's stated weakest_assumption: they highlight the risk that the correlation surface after stain-conditioned preprocessing is not peaked at the true transform; I see the dominant problem as the absence of any measurement at all, which makes that risk completely untested. The algorithm itself is plausible and deterministic, and the code is open, so I would not call the method fraudulent or incoherent; I would call the submission incomplete as evidence. A benchmark with landmark-based TRE is a concrete check that would settle whether the abstract's claim holds. If such results were added and acceptable, the paper could be reconsidered; as is, the rejection stands. Hence no verdict change, and only partial agreement with the reader's weakest_assumption because their rationale already flags the missing metrics but their stated weakest assumption is a technical condition inside the black box.","tokens_in":7376,"tokens_out":5934,"duration_ms":72232,"concrete_test":"Run the released STAR code on all ANHIR pairs with challenge-provided landmarks and on ACROBAT's validation subset. For each pair, apply the estimated rigid transform and compute the target registration error (TRE) at the provided landmark points in original slide coordinates; report median and 95th percentile TRE, and success rate (TRE below, e.g., 5% of image diagonal). Compare against at least one existing rigid baseline (e.g., ASHLAR or plain normalized cross-correlation) and record wall-clock time per pair on fixed hardware. If median TRE or success rate is not within an accepted range or no better than baseline, the abstract's 'stable/robust' claim is falsified; if it is, the empirical gap is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper is empirical: that STAR produces fast, stable, robust rigid alignments across multi-stain serial WSIs. The manuscript never tests this claim quantitatively. Section IV.A describes ANHIR and ACROBAT, and Section IV.B gives implementation details and a sentence that registration took 1–2 minutes per slide, but there are no landmark-based errors, no comparison with existing methods, no failure rates, and no timing benchmark with hardware specification. Section V.B offers qualitative 'visual inspections confirmed stable alignment,' which cannot support the abstract's robustness statement. Consequently, even if every equation and code path is correct, the paper's central contribution remains unverified: the method could be producing systematically wrong alignments on a substantial fraction of pairs. This is more load-bearing than the correlation-surface peakedness assumption, because that assumption is precisely what would be tested by a quantitative evaluation; currently the paper skips the test. The open-source release means the gap is fillable, but as submitted the claim overreaches.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents STAR, an open-source rigid registration framework for serial whole-slide histological images. The method combines stain-conditioned preprocessing (histogram equalization, inversion, thresholding, blur) with a hierarchical coarse-to-fine correlation search over in-plane rotation and translation, plus adaptive kernel scaling, quality control, and optional human-in-the-loop refinement. The authors claim that on the ANHIR 2019 and ACROBAT 2022 datasets STAR consistently produces stable alignments within minutes per slide, and they position it as a lightweight, reproducible baseline for multi-stain registration and AI training-data preparation. The manuscript includes a pipeline description, algorithm pseudocode, qualitative figures, and an open-source code release, but it contains no quantitative evaluation of registration accuracy, no comparison to existing methods, and only a terse statement about runtime.","tokens_in":7735,"tokens_out":3218,"duration_ms":40913,"significance":"If the claimed performance were substantiated, STAR would be a useful and practical contribution: a fast, simple rigid-registration baseline to complement deformable and deep-learning methods, with open-source code, batch processing support, and built-in quality control. The method is self-contained and does not appear to overfit to benchmark outcomes, which is a strength. However, the central empirical claim—robust, stable, fast alignment on ANHIR and ACROBAT—is currently unsupported by any quantitative evidence. The paper's value hinges on that claim, so the missing evaluation is the decisive issue.","major_comments":[{"comment":"The abstract states that STAR 'consistently produced stable alignments within minutes per slide' on ANHIR and ACROBAT, but the manuscript provides no quantitative results. Section IV only describes the datasets and implementation details; Section V.B reports 'visual inspections confirmed stable alignment,' which cannot support claims of robustness or consistency. There are no landmark-based target registration errors, no success/failure rates, no error bars, and no comparison with any baseline or existing rigid/affine registration method. Because the whole method reduces to an argmax over a correlation surface (Eq. 4), a systematic failure mode on a nontrivial fraction of pairs would be invisible in a few selected visual examples. The authors should add a full quantitative evaluation on ANHIR (which has ground-truth landmarks) and ACROBAT, including per-case error distributions, failure","section":"Abstract; §IV.A; §IV.B; §V.B"},{"comment":"The coarse-to-fine strategy assumes that the 10-degree angular stride and 10-pixel translation stride do not miss the global correlation peak. This assumption is load-bearing: if the coarse stage selects a wrong local mode, the fine stage (restricted to [θc−10, θc+10] and a ±50 px crop) cannot recover the true alignment. The manuscript offers no evidence that the correlation surface is sufficiently peaked or unimodal for the chosen grid spacing. The authors should provide a sensitivity analysis or a synthetic experiment with known ground-truth transforms to show that the coarse grid captures the global peak, or describe an adaptive scheme to avoid this risk.","section":"§III.D.1; Eq. (4)"},{"comment":"The only runtime statement, 'rigid registration per slide pair required approximately 1–2 minutes,' is not a reproducible benchmark. No hardware configuration (GPU model, CPU, memory), WSI dimensions, number of pairs, or per-case timing distribution is reported. Since 'fast' is a headline contribution, the authors should provide a proper timing table with hardware specifications, data sizes, and variance across cases, and ideally a comparison with a standard CPU-based registration baseline.","section":"§IV.B"},{"comment":"The stain-conditioned preprocessing pipeline—histogram equalization, intensity thresholds (<30 and <50), Gaussian blur, inversion, and adaptive kernel scaling—is central to the claimed cross-stain robustness, but none of these choices are validated or ablated. For example, the threshold at 30 for target images and the threshold at 50 for reference images are hard-coded, and no evidence shows they generalize across the stains and scanners represented in ANHIR/ACROBAT. An ablation study (e.g., removing inversion, blur, or each thresholding step) would substantiate the design and help users understand failure modes. This is particularly important because Section V.C already concedes that 'extreme staining artifacts or scanner-specific noise can reduce correlation responses,' so the preprocessing choices need quantitative support.","section":"§III.C; §III.D.3"}],"minor_comments":[{"comment":"Reference [2] contains a LaTeX artifact in the author list ('g.-i. family=Lotz, given=JM'), and reference [11] is missing complete bibliographic details. Please clean up the bibliography.","section":"References"},{"comment":"The inner product in Eq. (1) is not formally defined; since Φ_R and rotated/translated Φ_T may have different support, the correlation measure should be specified precisely (e.g., normalized cross-correlation or zero-normalized cross-correlation). This matters for understanding what the argmax actually optimizes.","section":"Eq. (1); §III.A"},{"comment":"Figure 2 shows representative cases but lacks scale bars, annotations of which panel corresponds to which dataset/stain, and any indication of alignment accuracy at high resolution. Adding landmark overlays or checkerboard/divider visualizations would strengthen the qualitative evidence.","section":"Fig. 2"},{"comment":"The ResNet-18 foreground detector is described as 'pre-trained' and 'modified with a binary classification head,' but no details are given about the training data, fine-tuning procedure, or expected foreground-mask accuracy. At minimum, state the source of the pretrained weights and whether the head was trained specifically for histology.","section":"§III.B"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations, but the abstract overstates the evidence: the ANHIR/ACROBAT evaluation is qualitative only, and the runtime claim is a single sentence. The missing quantitative evaluation is substantial but fixable within the manuscript's scope, so I recommend major revision rather than rejection. If the authors add proper landmark-based metrics, baselines, and hardware-specific timing, the paper could become a solid contribution to the registration toolkit literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the STAR pipeline is a sensible integration of standard building blocks—stain-conditioned preprocessing, hierarchical correlation, adaptive kernel scaling, quality control—and the authors released the code, which is genuinely useful for the community. But the paper's central claim, \"stable alignments within minutes on ANHIR/ACROBAT,\" has no quantitative support: no landmark errors, no baselines, no runtimes with hardware, no failure rates. The only evidence is visual inspection of a few cases. That's a load-bearing gap, not a minor omission.\n\nWhat the paper does well: the method is described clearly, the design choices are reasonable and mostly standard, and the authors are transparent about failure modes in Section V.C. The open-source release is a real contribution—a lightweight, reproducible rigid baseline is something the field lacks. The preprocessing asymmetry (reference H&E vs. target IHC/special stains) is well motivated, and the coarse-to-fine search with 10° then 1° increments is a sensible way to keep the search tractable.\n\nWhere it falls short: the evaluation section is essentially absent. The abstract promises robustness to cross-stain variability and partial tissue overlap, but the experiments never measure alignment accuracy. ANHIR and ACROBAT both provide landmarks; reporting TRE or at least a comparison with a simple baseline would be straightforward. The stress-test note is correct: this is more fundamental than the question of whether the correlation surface is peaked. If the correlation is unreliable, the method fails; if it is reliable, it works—but we need numbers to know which. The hand-picked free parameters (downsample factor, strides, thresholds) are not derived from the data, which is fine for a heuristic pipeline, but it makes the robustness claim even more dependent on empirical validation. The authors do note the partial-tissue-overlap handling in Section III.E, but no evidence is shown.\n\nOne minor thing: the runtime claim (1–2 minutes per slide) appears in the implementation details without hardware specs or a proper benchmark. That's exactly the kind of number that needs a table.\n\nWho this is for: someone building a rigid registration baseline for histology and wanting a starting point. The code may be worth trying directly. The paper itself, though, is not yet a credible evidence-based contribution.\n\nMy recommendation: I would not send this to peer review in its current form. The missing quantitative evaluation is a rejection-level issue, not a revision-level one. That said, the gap is fillable: the code is public, the datasets are public, and adding landmark-based errors and baseline comparisons would likely make this a solid submission. Revise, add the numbers, then resubmit.","headline":"STAR is a clear, well-written, open-source rigid registration pipeline for histology, but the paper's central accuracy claims are unsupported by any quantitative evaluation.","tokens_in":8087,"tokens_out":2383,"would_cite":false,"duration_ms":28169,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lightweight rigid pipeline aligns multi-stain whole-slide histology images in minutes.","keywords":["histopathology","whole-slide image registration","rigid registration","multi-stain alignment","template matching","immunohistochemistry","cross-stain preprocessing","computational pathology"],"falsifier":"Run STAR on a set of serial H&E-IHC slide pairs and compare its estimated rotation and translation with manually placed landmarks on the 32x thumbnails; if a non-negligible fraction of pairs has rotational error of 10 degrees or more, or translational error of 10 pixels or more at that scale, the coarse-to-fine peak assumption is falsified.","tokens_in":7331,"feed_emoji":"🔬","tokens_out":4840,"duration_ms":55341,"temperature":0.7,"pith_summary":"This paper argues that for consecutive serial sections, rigid alignment is often sufficient, and that a deliberately simple pipeline built from stain-conditioned preprocessing plus hierarchical correlation can replace heavier deformable or deep-learning tools for multi-stain whole-slide registration. It introduces STAR, which registers all target stains to a single reference by maximizing a correlation score over rotation and translation on downsampled thumbnails, then refines the result. Evaluated on the ANHIR 2019 and ACROBAT 2022 datasets, the claim is stable alignment across H&E, special histochemical stains, and IHC markers, within minutes per slide, despite cross-stain variability and partial tissue overlap. If correct, STAR gives pathologists and AI-data builders a reproducible lightweight baseline instead of requiring complex deformable models.","feed_headline":"STAR aligns multi-stain histology slides in minutes","feed_subtitle":"Open-source correlation pipeline pairs H&E with special stains and IHC without heavy deep learning.","key_machinery":"The mechanism is the correlation-maximization registration objective: for each target, STAR selects the rotation and translation that maximize the inner product between the processed reference and the transformed target. It implements this as a coarse-to-fine convolution search over a GPU rotation bank, with an adaptive kernel-scaling rule that downsamples the template when it is larger than the cropped target. The load-bearing part is the stain-conditioned preprocessing: it is what makes the correlation surface peak at the true rigid transform despite different stains.","core_discovery":"The central claim is that a multi-stage template-correlation search, after stain-aware preprocessing, is sufficient to align serial whole-slide images across very different staining protocols. The objective is the arg max over in-plane rotation and translation of the correlation between a processed grayscale reference and each target. STAR builds a bank of 36 rotated reference templates at 10-degree steps, correlates each against the processed target on a 32x downsampled thumbnail, selects the best peak, then refines around it at 1-degree and pixel-level strides. The key design choice is asymmetric preprocessing: reference sections (usually H&E) are histogram-equalized, inverted, and contras","pith_inferences":["If the preprocessing truly normalizes stain appearance, the same correlation objective could plausibly extend to other conserved-architecture modalities such as multiplex immunofluorescence or cytology smears; the paper only mentions this as future work.","A testable extension is to measure how far the coarse 10-degree and 10-pixel grid can be coarsened before the capture rate drops, by perturbing known transforms and running the pipeline on synthetic rotations and shifts.","The implicit claim that rigid alignment suffices for consecutive sections likely degrades as section spacing grows; one could test failure rate as a function of cutting distance, since finer structural correspondence will worsen even with perfect preprocessing."],"forward_implications":["Consecutive-section workflows can adopt a rigid baseline that runs in roughly one to two minutes per slide pair, making large-scale paired dataset construction practical.","H&E-IHC and multi-IHC panels can be aligned to a single reference, enabling virtual staining and biomarker-prediction training data without specialized multiplex imaging.","The open-source implementation gives other researchers a reproducible comparison point; affine or deformable refinements can be appended later without replacing the rigid core.","The exported registration parameters and tissue masks allow downstream patch extraction to be restricted to corresponding tissue regions."],"supporting_citations":[{"why":"Supplies the premise that rigid alignment often suffices for consecutive sections, motivating the entire approach.","marker":"[8]"},{"why":"Represents the existing open-source WSI registration workflow that STAR positions itself against as a lighter rigid baseline.","marker":"[11]"},{"why":"Provides the ANHIR challenge dataset and landmark ground truth used to evaluate STAR across organs and stains.","marker":"[12]"},{"why":"Provides the ACROBAT multi-institution dataset with H&E, IHC, and special stains used to test clinical realism and stain diversity.","marker":"[13]"}],"fun_headline_variants":["STAR aligns multi-stain slides in minutes, no deep learning needed","Fast rigid alignment for histology: STAR does it in minutes","STAR: open-source rigid registration for H&E and IHC","Align cross-stain slides fast with STAR's correlation approach","Minutes to align serial histology: meet STAR"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything rests on the coarse 10-degree, 10-pixel correlation search landing close enough to the true alignment after the stain-normalizing preprocessing; if staining differences flatten or shift the correlation peak beyond that grid, the claimed cross-stain stability fails.","fun_headline_variants_meta":{"raw":{"variants":["STAR aligns multi-stain slides in minutes, no deep learning needed","Fast rigid alignment for histology: STAR does it in minutes","STAR: open-source rigid registration for H&E and IHC","Align cross-stain slides fast with STAR's correlation approach","Minutes to align serial histology: meet STAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000537,"raw_usage":{"total_tokens":2438,"prompt_tokens":792,"completion_tokens":1646,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1561}},"tokens_in":536,"tokens_out":1646,"duration_ms":14107,"temperature":1.0,"reasoning_tokens":1561,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:14:22.478044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run STAR on a set of serial H&E-IHC slide pairs and compare its estimated rotation and translation with manually placed landmarks on the 32x thumbnails; if a non-negligible fraction of pairs has rotational error of 10 degrees or more, or translational error of 10 pixels or more at that scale, the coarse-to-fine peak assumption is falsified.","supporting_citations":[{"cited_title":"Stitching and registering highly multiplexed whole-slide images of tissues and tumors using ashlar,","cited_arxiv_id":null,"evidence_quote":"Supplies the premise that rigid alignment often suffices for consecutive sections, motivating the entire approach."},{"cited_title":"An open-source whole slide image registration workflow at cellular precision using Fiji, QuPath and Elastix,","cited_arxiv_id":null,"evidence_quote":"Represents the existing open-source WSI registration workflow that STAR positions itself against as a lighter rigid baseline."},{"cited_title":"ANHIR: Automatic Non-Rigid Histological Image Registration Challenge,","cited_arxiv_id":null,"evidence_quote":"Provides the ANHIR challenge dataset and landmark ground truth used to evaluate STAR across organs and stains."},{"cited_title":"The ACROBAT 2022 challenge: Automatic registration of breast cancer tissue,","cited_arxiv_id":null,"evidence_quote":"Provides the ACROBAT multi-institution dataset with H&E, IHC, and special stains used to test clinical realism and stain diversity."}],"review_version":1}