{"id":"b99e13bb-8c10-478e-ac85-f1185f07deba","arxiv_id":"2511.03826","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"CORE is a coarse-to-fine whole-slide image registration pipeline that aligns H&E, IHC, PAS and mIF slides at tissue and nuclei level, reporting competitive-to-better landmark errors than published baselines.","lead":"CORE is a two-stage pipeline that registers multi-stained whole-slide pathology images: it first aligns tissue masks and deep features, then refines alignment using nuclei centroids with shape-aware point-set registration and coherent point drift. On three public and two private datasets it reports lower landmark error than several published baselines, with coarse alignment in about 12–20 seconds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fine-stage claim of nuclei-level correspondence is not supported by any nuclei-level evaluation metric; all quantitative results use sparse tissue landmarks.","rationale":"The reader's weakest assumption focuses on the input side of the fine stage: the morphology-based nuclei detector must return complete, mutually corresponding centroid sets. This is a valid and acknowledged limitation. However, the stress-test pass identifies a more directly load-bearing gap on the output side: even if the detector works perfectly, the paper never measures whether individual nuclei are actually brought into correspondence. The central claim explicitly asserts 'nuclei-level correspondence,' but the reported TRE/rTRE metrics are computed on sparse tissue landmarks, which may not sit on nuclei. Thus, the claim is not falsifiable from the presented evidence. This is an epistemic gap, not an algorithmic flaw, and it is more fundamental because it affects the interpretation of all fine-stage results. The proposed concrete test is feasible: the HYRECO re-stained subset is a same-section experiment where ground-truth nucleus correspondence is well-defined, and the authors already possess the registration pipeline and detector. If the test passes, it would strengthen the claim; if it fails, the conditional verdict should escalate. Partial agreement with the reader: their detector concern is related to the fine stage, but my concern is distinct and, in my view, more load-bearing for the stated contribution.","tokens_in":19029,"tokens_out":10991,"duration_ms":105905,"concrete_test":"On the HYRECO re-stained subset (54 H&E–PHH3 pairs, same physical section), run the full CORE pipeline (or at least the fine stage). After registration, detect nuclei in both target and registered source using the paper's own morphology-based detector (or a validated Hover-Net model). Match the two centroid sets via Hungarian matching with a maximum distance threshold of 10 μm (approximately the reported fine-stage median TRE converted to pixels at 0.24 μm/px). Report the match rate (fraction of target nuclei with a matched source centroid within threshold) and median matching distance. Compare these nuclei-level metrics with the coarse-stage results and with a random-alignment baseline. If the fine-stage match rate is not substantially higher than the coarse-stage rate, the claim of 'precise nuclei-level correspondence' is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract claims 'accurate nuclei-level registration' and 'precise nuclei-level correspondence across modalities', and Section 3.2 describes the fine stage as delivering cellular-level precision. However, Tables 2–6 evaluate performance exclusively with landmark-based TRE/rTRE, where landmarks are manually placed anatomical points (e.g., 11–19 per HYRECO section, ~43 per re-stained pair). These landmarks are not shown to coincide with nuclei centroids. A low TRE after fine registration demonstrates tissue-level alignment but does not establish that individual nuclei correspond between modalities. In fact, the fine stage optimizes a point-set objective on nuclei centroids, so the evaluation metric should directly assess nucleus-to-nucleus matching. Without a nuclei-level metric, the central novelty remains unverified. Additionally, on the ACROBAT dataset (the largest public benchmark), fine registration was not performed at all (slides are 10×), so the generalisability of the nuclei-level claim is untested on that cohort. This concern is distinct from the acknowledged detector-dependence limitation: even a perfect detector would not rescue the claim if the evaluation cannot resolve whether nuclei are aligned. The paper's qualitative overlays (Figs. 4–9) are suggestive but do not substitute for quantitative cellular correspondence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CORE, a two-stage whole-slide image (WSI) registration framework. The coarse stage combines prompt-based tissue-mask extraction (Florence-2-SAM), a morphology-based rigid alignment module (TriMorph), XFeat dense feature matching, and hierarchical non-rigid NCC optimization with smoothness regularization. The fine stage detects nuclei centroids via a morphology/watershed pipeline, performs a shape-aware point-set rigid alignment (Eqs. 4–7, Powell optimization), and then estimates local non-rigid deformations with Coherent Point Drift (CPD). The method is evaluated on ACROBAT, ANHIR, HYRECO, Multi-IHC CRC, and REACTIV AS, using TRE/rTRE-based landmarks and runtime comparisons, and the authors claim that CORE outperforms state-of-the-art methods in precision, robustness, and generalisability.","tokens_in":19386,"tokens_out":5953,"duration_ms":55854,"significance":"If the reported claims are substantiated, CORE would be a practically valuable contribution: the coarse stage is fast (e.g., 14.37 s on ACROBAT versus 50–120 s for published baselines), the method is tested on five cohorts spanning bright-field and immunofluorescence modalities, and open-source code plus a TiaViz demo are provided. The use of a classical, training-free nuclei detector is also attractive for data-scarce settings. However, the headline claim of 'accurate nuclei-level registration' is not directly supported by the current evaluation, which relies entirely on tissue-level landmark TRE/rTRE. The coarse-registration evidence is more credible than the fine-stage evidence; the fine-stage and 'generalises without manual tuning' claims need additional support before the paper can be accepted.","major_comments":[{"comment":"The abstract and §3.2 claim 'precise nuclei-level correspondence across modalities' and 'cellular-level precision.' But all quantitative results in Tables 2–6 use landmark-based TRE/rTRE (§4.2), and the landmarks (e.g., 11–19 per HYRECO section, ~43 per re-stained pair) are manually placed anatomical points, not shown to coincide with nuclei centroids. A low TRE after fine registration demonstrates tissue-level alignment, not nucleus-to-nucleus correspondence. Since the fine stage is a point-set registration on nuclei centroids, the evaluation should include a direct nuclei-level metric, such as mutual-nearest-neighbour centroid precision/recall, distances between corresponding nuclei, or a cell-matching F1 score. Without such a metric, the paper's central novelty is unverified. This is separate from the acknowledged detector-dependence limitation in §4.4.","section":"§3.2, Tables 2–6"},{"comment":"Table 3 reports 'Fine CORE' as having the lowest AArTRE and AMrTRE on ANHIR, but §4.3 states that fine-shape-aware registration was performed only on slides at 20×/40× magnification, while lower-resolution samples (1.25×–10×) were excluded. The baselines (DFBR, DeeperHistReg, HistokatFusion, etc.) are evaluated on the full ANHIR set. Thus the fine-stage comparison is not head-to-head: the improved global metrics could reflect case selection rather than method superiority. Please report per-subset results for CORE and all baselines on the same 20×/40× pairs, or provide a full-set fine-registration result if feasible. This is necessary to support the 'outperforms at fine registration' claim.","section":"Table 3"},{"comment":"No confidence intervals, standard errors, or significance tests are reported for any metric. For example, in Table 2, CORE's AMTRE90 of 139.00 µm is only 2.6 µm below HistokatFusion (141.64) and 1.3 µm below DeeperHistReg (140.33). Median/mean differences of this size may be within case-level variability. Because the abstract claims 'outperforms current state-of-the-art methods' across five datasets, the paper should provide paired statistical comparisons (e.g., bootstrap confidence intervals or Wilcoxon signed-rank tests over WSI pairs) and report the number of pairs contributing to each aggregate. The same applies to runtime comparisons, where only single means are given.","section":"Tables 2–6"},{"comment":"The introduction says the framework 'generalises across staining protocols without manual tuning,' but §3.1.1 states that gamma correction ranges (1.0–1.2 for H&E, 0.4–0.8 for IHC) were 'empirically determined based on evaluations conducted across the datasets,' and many other parameters are fixed values with no sensitivity analysis: Dice threshold 0.7, rotation step 10°, XFeat precision/iteration settings, Powell tolerance, CPD regularization parameters, and shape weight γ. Please clarify explicitly whether these were fixed a priori across all datasets or tuned per dataset. If they were tuned, the 'without manual tuning' claim should be softened; at minimum, provide a sensitivity study for the most influential parameters (e.g., γ in Eq. (5) and CPD α, β in Table A.4).","section":"§3.1.1–§3.1.2, Tables A.2–A.4"}],"minor_comments":[{"comment":"The REACTIV AS dataset reference appears as '[citation]' in the text; it should be [33].","section":"§4.1.5"},{"comment":"§4.1.1 states ACROBAT contains 3,406 WSIs, while Table 1 reports 4,212 WSIs. Please reconcile these numbers.","section":"Table 1 vs §4.1.1"},{"comment":"Figure 6 caption says source (PAS), but HYRECO stain sets are H&E, PHH3, CD8, CD45; the text describes PHH3. Please correct the caption or clarify the stain.","section":"Figure 6 caption"},{"comment":"Report the number of image pairs used for each metric, especially for the private REACTIV AS dataset (11 paired samples) and the HYRECO subsets, so that the aggregate values can be interpreted.","section":"Tables 4 and 6"},{"comment":"A few in-text citation numbers appear inconsistent with the reference list (e.g., 'Huang et al. [13]' while [13] lists Wei et al.). Please verify all citation labels.","section":"§2 Related Works"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of intentional circularity or citation manipulation; the main shortcomings are in evaluation design, not in the methodology itself. The paper is within scope for q-bio.QM, and the coarse-stage work is promising. If the authors add a direct nuclei-level evaluation, make the ANHIR fine-stage comparison head-to-head, and provide statistical support for the superiority claims, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read CORE carefully. The stress-test note lands, and the reader's conditional verdict is about right.\n\nThe headline claim is that CORE achieves precise nuclei-level correspondence across H&E, IHC, PAS, and mIF. But every quantitative accuracy number in Tables 2–6 is TRE or rTRE on manually placed landmarks. Those landmarks are not shown to coincide with nuclei centroids, and in HYRECO they are 11–19 per section. So the fine stage's actual objective—aligning nuclei point sets—is never evaluated with a nuclei-level metric. Qualitative overlays don't fix that. This is not a completeness nit; it is the main advertised novelty. The paper's own limitations paragraph admits the fine stage depends on accurate nuclei detection, which is honest, but even a perfect detector would not rescue the claim when the metric cannot resolve whether nuclei are aligned.\n\nCredit where due. The coarse stage is fast, about 14 seconds on ACROBAT, and competitive with published methods across several datasets. Prompt-based tissue masking with Florence2-SAM followed by XFeat is a reasonable integration, not a breakthrough but a working one. The fine-stage design—morphology watershed centroids, a hybrid spatial-shape distance, Powell rigid alignment, CPD non-rigid—is coherent and described well enough to reproduce. The stepwise tables let you see error drop at each stage, which is honest engineering reporting. The self-citation of DFBR and TIATOOLBOX is minor and not circular.\n\nSoft spots, in proportion. The ANHIR fine-stage comparison covers only the 20x/40x subset while the baselines cover all pairs, so the \"outperforms\" line is not a head-to-head. There are no confidence intervals, significance tests, or per-pair variance figures anywhere; with five datasets and several hyperparameters tuned on them, that is material. Gamma correction range, Dice threshold, rotation step, and XFeat settings are empirically determined and not given sensitivity analysis. The fine stage runs for hundreds of seconds, so the efficiency advantage is really a coarse-stage advantage. These are fixable, not fatal.\n\nWho gets value: anyone working on multi-stain spatial pathology or WSI registration who wants a practical baseline to compare against. It is a systems paper with real engineering content, a broad benchmark footprint, and promised open code. It needs revision before acceptance—an evaluation metric that actually reports nucleus-to-nucleus correspondence, matched subsets, and some statistical reporting—but it deserves a serious referee. I would send it to peer review.","headline":"Useful coarse-to-fine WSI registration platform with a genuinely fast coarse stage, but the nuclei-level claim outruns the evidence: all fine-stage numbers are TRE on sparse manual landmarks, not on nuclei.","tokens_in":19874,"tokens_out":2139,"would_cite":true,"duration_ms":23318,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage pipeline aligns multi-stain tissue slides down to nuclei-level precision.","keywords":["whole slide image registration","multi-stain alignment","nuclei point-set registration","coarse-to-fine","coherent point drift","shape-aware registration","computational pathology","immunofluorescence"],"falsifier":"Measure CORE's fine-stage target registration error on slide pairs with artificially withheld or poorly detected nuclei—for example, by running the watershed detector on patches with dense overlapping lymphocyte clusters and comparing against manual landmarks. If the median fine error rises sharply when detection completeness drops below roughly 90 percent, the central nuclei-correspondence premise is falsified.","tokens_in":18961,"feed_emoji":"🔬","tokens_out":5851,"duration_ms":51353,"temperature":0.7,"pith_summary":"The paper introduces CORE, a coarse-to-fine whole-slide image registration engine intended to align tissue sections stained with different protocols—H&E, IHC, PAS, and multiplex immunofluorescence—down to nuclei-level precision. Its central claim is that a fast coarse alignment based on tissue masks and deep local features, followed by a fine alignment based on automatically detected nuclei centroids and deformable point-set matching, beats existing registration methods on accuracy, generality, and speed. The payoff, if the claim holds, is that researchers can combine information across stains without manual landmarking or stain-specific tuning, enabling cell-level multimodal analysis of large pathology images. The paper evaluates this on three public and two private cohorts, reporting median fine-stage errors below one micrometre on re-stained sections and a coarse stage that completes in roughly 15–20 seconds.","feed_headline":"Two-stage CORE pipeline aligns multi-stain slides to nuclei precision","feed_subtitle":"Coarse stage runs in ~15–20 s; fine nuclei point-set alignment cuts error below 1 µm on re-stained slides.","key_machinery":"The load-bearing object is the automatically detected nuclei centroid point set. The method's distinctive identity is a hybrid shape-aware distance metric that combines Euclidean distance between transformed source nuclei and target nuclei with the absolute difference of their normalised per-nucleus areas, balanced by a weight parameter set to 0.3. This metric drives a derivative-free Powell optimisation for rigid alignment, and the resulting correspondences feed Coherent Point Drift, a probabilistic non-rigid point-set registration algorithm, to produce the final dense displacement field. The paper also relies on a prompt-based tissue mask extraction step and an accelerated deep feature mat","core_discovery":"On the paper's own terms, the central discovery is that whole-slide registration across very different stains can be made both fast and nuclei-accurate by separating the problem into a morphology-first coarse stage and a nuclei-point-set fine stage. The coarse stage uses prompt-based tissue masks, centre-of-mass alignment, discrete rotation search, and accelerated dense feature matching at low magnification to produce a global alignment in about 15–20 seconds. The fine stage then detects nuclei centroids in the target and coarsely registered source, applies a shape-aware rigid point-set alignment that mixes spatial distance with normalised nuclear area, and finishes with Coherent Point Drift","pith_inferences":["The paper leaves implicit that the coarse stage's speed makes it suitable for interactive browsing or triage, but the reported runtime is for the alignment computation itself, not for full-resolution rendering or visualisation.","A natural stress test the paper does not run is registration of slide pairs with large tissue folds or tears inside the tissue mask; such cases would exercise the rigid coarse estimate and the Jacobian-based folding correction simultaneously.","The point-set machinery could be transferred to other intrinsic landmarks—glands, vessels, or immune-cell clusters—which would extend the method to tissues where nuclei are not the most reliable modality-invariant feature.","The explicit dependence on nuclei detection suggests that a version of CORE using a learned detector, trained only where annotated nuclei exist, could improve the fine stage in dense tissue while keeping the coarse-to-fine structure intact."],"forward_implications":["A fast, accurate coarse alignment becomes available as a standalone step, making interactive or high-throughput registration possible where nuclei-level precision is not required.","The same untuned pipeline can align bright-field stains such as H&E, IHC, and PAS with multiplex immunofluorescence, a cross-modality combination that is difficult for stain-specific methods.","Cell-level downstream analyses, such as matching single cells across stains or linking phenotype to morphology, can be run on the fine-stage output without manual landmark selection.","Because nuclei detection is morphology-based rather than trained per stain, the method transfers to new stain protocols without needing annotated nuclei datasets.","If the reported sub-micrometre errors on re-stained sections are reproducible, the fine stage reaches the precision needed to treat nuclei as corresponding biological units across staining rounds."],"fun_headline_variants":["Two-stage CORE cuts multi-stain alignment error below 1 µm","CORE aligns multi-stain slides with nuclei-level precision in seconds","Fast coarse-to-fine WSI registration hits sub-micron nuclei accuracy","Nuclei-point-set registration makes multi-stain WSI alignment spot-on"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fine stage assumes that automatically detected nuclei in one stain line up as corresponding landmarks with nuclei in the other stain; if detection fails or produces non-corresponding centroids in dense, overlapping, or artefact-heavy tissue, the point-set fit will target the wrong correspondences.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage CORE cuts multi-stain alignment error below 1 µm","CORE aligns multi-stain slides with nuclei-level precision in seconds","Fast coarse-to-fine WSI registration hits sub-micron nuclei accuracy","Nuclei-point-set registration makes multi-stain WSI alignment spot-on"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":2866,"prompt_tokens":738,"completion_tokens":2128,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2048}},"tokens_in":482,"tokens_out":2128,"duration_ms":13306,"temperature":1.0,"reasoning_tokens":2048,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:49:36.793946+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure CORE's fine-stage target registration error on slide pairs with artificially withheld or poorly detected nuclei—for example, by running the watershed detector on patches with dense overlapping lymphocyte clusters and comparing against manual landmarks. If the median fine error rises sharply when detection completeness drops below roughly 90 percent, the central nuclei-correspondence premise is falsified.","supporting_citations":[],"review_version":1}