{"id":"fe553b78-586e-4eb8-a64e-1836a3514975","arxiv_id":"2508.04450","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single-center whole-body CT registration model with region-specific deformation blocks matches or beats a foundation model on abdominal organs and generalizes to external datasets without fine-tuning.","lead":"This paper describes TotalRegistrator, a lightweight deep learning model that aligns whole-body CT scans from the same patient taken at different times, using standard UNets and a region-by-region deformation strategy. It reports competitive accuracy against a larger foundation model and a classical algorithm while training on about 11GB of GPU memory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential patient-level data leakage in the in-house train/test split could inflate reported performance gains.","rationale":"The reader's weakest assumption focused on the accuracy of TotalSegmentator-generated training masks and the diversity of a single-center dataset. While that is a legitimate concern, the paper already uses manually refined masks for the test set, which mitigates evaluation bias, and the external datasets provide some evidence of generalizability. The more load-bearing issue is the potential lack of a patient-disjoint split in the in-house experiment. If patient overlap exists, the primary result—outperforming uniGradICON and Elastix on RUMC—is not trustworthy, regardless of mask quality. The paper's silence on the splitting method makes this a concrete, addressable methodological gap. Verifying patient-level separation is a standard expectation in medical imaging studies and would settle the concern. The reader's CONDITIONAL verdict already reflects the need for code/weights and official test data, so adding this condition does not change the overall verdict level, but it should be an explicit requirement before acceptance. Therefore, I maintain CONDITIONAL (UNCHANGED) but with an added condition on demonstrating a patient-disjoint split.","tokens_in":10575,"tokens_out":5942,"duration_ms":69817,"concrete_test":"Obtain the patient identifiers for the 591 training pairs and 104 evaluation pairs from the RUMC dataset and compute whether any patient appears in both splits. If overlap exists, rerun the in-house evaluation using a strictly patient-disjoint split (e.g., by randomly assigning all pairs from each patient to either training or test, then retraining and re-evaluating). If the average DSC gap between TotalRegistrator and the baselines shrinks substantially or reverses, the original results are inflated by leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TotalRegistrator generally surpasses Elastix and uniGradICON on the in-house dataset rests on the integrity of the training/test split. In Section 3.1, the RUMC dataset is described as 695 pairs of longitudinal CT scans from individual patients, but the paper does not state that the 591 training pairs and 104 evaluation pairs (Table 1) are patient-disjoint. If the split was performed at the pair level rather than the patient level, the same patient could appear in both training and test sets, allowing the model to memorize patient-specific anatomical patterns and inflate the reported Dice scores. This would bias the primary comparison against baselines, which were not trained on any RUMC data. The problem is amplified because the model heavily uses a segmentation overlay loss that can align to patient-specific organ shapes if seen during training. While the paper mentions careful curation for demographic and scanner diversity, it omits any patient-level split guarantee. This is a more concrete and potentially more severe threat to the main claim than the quality of automatically generated training labels, because it directly undermines the validity of the primary evaluation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TotalRegistrator is a learning-based whole-body CT registration method trained on a single in-house longitudinal dataset. The architecture uses standard UNet blocks for affine and deformable stages, plus a field-decomposition strategy that divides the deformation field into thoracic, abdominal, bone, and whole-body components; components are trained independently and then composed. The model is evaluated on 104 in-house whole-body pairs (intra- and inter-patient) and on three external datasets (AbdomenCTCT, NLST, H108M). The central claim is that TotalRegistrator generally outperforms Elastix and uniGradICON on multi-organ abdominal Dice while remaining competitive on external data, and that it can be trained on 11GB GPU memory. The authors acknowledge limitations in bone registration, folding in small organs, and the segmentation-mask-driven loss.","tokens_in":10819,"tokens_out":6633,"duration_ms":72762,"significance":"If the reported results hold, the paper makes a useful practical contribution: it demonstrates that a lightweight, single-dataset-trained model with region-specific field decomposition can approach, and on abdominal structures exceed, a larger foundation model, lowering hardware barriers for multi-organ CT registration. The manuscript has concrete strengths: a carefully curated 695-pair whole-body dataset with demographic and scanner diversity, external evaluation, ablations of the individual registration blocks, a transparent limitation section, and a plan for public code. The main caveats are the ambiguous train/test patient split, unvalidated automatic training labels, and lack of statistical significance testing, which currently leave the strength of the central comparative claim uncertain.","major_comments":[{"comment":"The paper must state explicitly that the 591 training and 104 evaluation pairs are patient-disjoint. The current text says 695 pairs 'from patients' and describes patient demographics, which suggests one pair per patient, but it never says 'one pair per patient' or that the split was stratified by patient. If the same patient contributed to both training and test sets, the segmentation-overlay loss could memorize patient-specific anatomy and inflate the reported DSC gain over Elastix and uniGradICON, which were not trained on RUMC data. This is load-bearing for the primary in-house claim. Please clarify, and if the split was not patient-disjoint, re-run the evaluation on a patient-disjoint split.","section":"Section 3.1, Table 1"},{"comment":"Training masks are produced by TotalSegmentator without expert refinement, while only test masks are manually refined. Because the segmentation-overlay loss is a core training signal, systematic TotalSegmentator errors on small structures such as the pancreas and gallbladder could bias the learned deformation fields and downstream organ-specific comparisons. The authors should quantify label noise (e.g., Dice between TotalSegmentator and refined masks on the test set) or otherwise justify that the automatic training labels are accurate enough to drive the reported results.","section":"Section 3.2(a), Fig. 6"},{"comment":"The paper uses phrases such as 'significantly better' and 'significant difference' (e.g., Section 4.1) without reporting statistical tests. Given the large variability for small abdominal organs shown in Fig. 7, the claimed superiority over Elastix and uniGradICON should be supported by paired tests or confidence intervals across the 104 in-house pairs. Without this, the central comparative claim is not fully supported.","section":"Section 4, Tables 2-5"}],"minor_comments":[{"comment":"External evaluations use the Learn2Reg validation subsets, not the official test sets. The abstract and Section 5.2 should consistently say 'validation subset' rather than phrasing that implies full official evaluation.","section":"Section 3.1(b), Abstract"},{"comment":"The heading contains a typo: 'Apporach' should be 'Approach'. Also, the term 'novel field decomposition' should be positioned against prior cascaded registration works (Zhao et al., 2019; Pham et al., 2024) to clarify the incremental contribution.","section":"Section 2, Fig. 4"},{"comment":"The list of annotated structures totals 12 if kidneys and lungs are counted as single labels, not the stated 13. Please clarify whether left/right organs are counted separately or whether the whole-body mask is included in the count.","section":"Section 3.2(a)"},{"comment":"The boxplots are dense; method names and organ labels are difficult to read at print size. Consider splitting by region or increasing font size.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the patient-split ambiguity is the main risk. The wording suggests one pair per patient, but the manuscript does not explicitly rule out pair-level splitting, and the central claim depends on this condition. If the split is already patient-disjoint, the paper is likely acceptable after revisions; I nonetheless recommend major revision because the current text does not let a reader verify a load-bearing condition. The external validation on challenge validation subsets is a known limitation, not a novelty concern. No suspicious citation pattern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-scoped empirical paper on lightweight whole-body CT registration. The field decomposition with independently trained region blocks plus a whole-body integration block is a sensible extension of cascaded registration, and the per-organ backward pass memory trick is a real practical contribution. The central claim—that you can train a competitive multi-organ model from scratch on a single curated longitudinal dataset with ~11GB VRAM—holds up for what is reported.\n\nWhat the paper does well: the ablations actually justify the decomposition, showing the standalone whole-body block is outperformed by region-specific blocks in their own domains. External evaluation on Learn2Reg validation subsets and the H108M liver dataset gives genuine out-of-distribution evidence, and the limitations section is unusually honest: bone registration is unstable, gallbladder/stomach folding persists, and the segmentation-driven loss biases toward global shapes.\n\nSoft spots: the biggest issue is an omitted detail rather than a demonstrated flaw. The paper never states that the 591 training pairs and 104 evaluation pairs from RUMC are patient-disjoint. Since the dataset is described as intra-patient longitudinal pairs, it is possible (though not stated) that some patients contribute multiple pairs. If the split was done at pair level rather than patient level, the segmentation-overlay loss could memorize patient-specific anatomy and inflate the in-house Dice against baselines that never saw any RUMC data. This needs a one-sentence clarification. Second, the code and weights are not yet released and the training data are private, so the headline numbers cannot yet be independently checked. Third, external results on AbdomenCTCT and NLST use validation subsets, which is consistent with how uniGradICON was evaluated but is still weaker than official test sets.\n\nThe citation pattern is fine; prior work is credited, including the authors' own SPIE preliminary version, which is a good model for incremental publication. I did not find any load-bearing fitting-as-prediction issue or internal contradiction.\n\nThis paper is for researchers in medical image registration, especially those who need a practical multi-organ baseline or want to train on a single hospital dataset. It deserves a serious referee. I would send it out with the patient-level split question as the first thing to resolve, and I would want a commitment to release code and weights before final acceptance.","headline":"A credible, useful whole-body CT registration model with a genuinely practical training trick; the main open question is whether the in-house train/eval split is patient-disjoint.","tokens_in":11363,"tokens_out":1681,"would_cite":true,"duration_ms":20472,"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 whole-body CT registration model trained on 591 pairs competes with specialist models on unseen tasks.","keywords":["image registration","whole-body CT","multi-organ registration","field decomposition","unsupervised learning","UNet","medical image foundation model","longitudinal CT"],"falsifier":"Manually correct the training-time organ masks for a subset of cases, retrain under identical settings, and compare held-out Dice; if scores change materially, the automatic labels were load-bearing. Alternatively, test on whole-body CT scans from scanners and patient populations not represented in the training cohort; a large drop in abdominal Dice would falsify the generalizability claim.","tokens_in":10485,"feed_emoji":"🩻","tokens_out":5888,"duration_ms":62357,"temperature":0.7,"pith_summary":"TotalRegistrator claims that a general-purpose whole-body CT registration model can be trained from scratch on one carefully curated longitudinal dataset of 591 whole-body scan pairs, without aggregating many public datasets or using heavy compute. The paper's key move is to replace one complex global deformation field with several independently trained region-specific fields, composed sequentially and finally smoothed by a whole-body block. On its held-out in-house test set, the model generally outperforms a classical iterative tool and a recent foundation model on multi-organ abdominal alignment, while lagging slightly on lungs and bone. On three external datasets it remains competitive with single-organ specialist models despite never being fine-tuned. If this holds, whole-body registration becomes feasible for hospitals with ordinary GPUs and modest training data.","feed_headline":"Whole-body CT registration rivals single-organ models","feed_subtitle":"A UNet model trained on 591 scan pairs matches specialist results on unseen organs without fine-tuning.","key_machinery":"The central mechanism is field decomposition: the whole-body deformation field is split into region-specific displacement fields for bone, thorax, and abdomen, each produced by an independent standard UNet after a shared affine alignment. The fields are composed sequentially, and a whole-body block, trained on the frozen outputs of the earlier blocks, integrates and regularizes them. Training memory is kept to about 11GB by performing forward and backward passes per organ instead of maintaining one cumulative gradient graph for all organs.","core_discovery":"The central claim is that a single well-curated longitudinal dataset of whole-body CT pairs is enough to train a general multi-organ registrator from scratch, provided the deformation problem is decomposed by anatomical region. The model uses standard UNet blocks, an unsupervised training loss combining mutual-information similarity, overlap of automatically generated organ masks, and deformation regularization. The authors report that, on their held-out in-house pairs, TotalRegistrator achieves the highest average Dice for abdominal organs among the compared methods in both intra-patient and inter-patient settings, with a lower folding rate than a recent foundation model in intra-patient ca","pith_inferences":["The design implies the decomposition, not the UNet itself, is the active ingredient; a testable extension is to add more region blocks (for example, head or neck) and check whether the whole-body integration block absorbs them without retraining earlier blocks.","Because only test labels were manually refined, the reported Dice may depend on the quality of automatically generated training masks; retraining with expert-corrected masks would reveal how much the automatic labels matter.","The per-organ backpropagation trick decouples memory use from the number of organs, so it could be transplanted to higher-capacity architectures such as transformer-based registration networks.","The reliance on segmentation masks in training means organs without reliable automatic masks are currently uncovered; the stated path to a full foundation model is to extend to all structures an automatic whole-body segmenter can label."],"forward_implications":["If the claim is correct, a whole-body multi-organ registration model can be built from a single well-curated longitudinal dataset rather than a large mixture of public datasets.","A standard UNet backbone plus the decomposition strategy is enough to compete with specialized single-organ models on unseen tasks, reducing the need for task-specific training.","Training on 11GB of GPU memory makes the approach practical for hospitals and smaller research groups.","The model's robustness to contrast variation, shown on multiphase liver data, suggests it can be applied to non-contrast and multiphase abdominal CT without retraining.","Bone and lung alignment remain weaker points, indicating that decomposition alone does not automatically make every anatomical region a specialist."],"supporting_citations":[{"why":"Supplies the automatic whole-body organ segmentations used as training supervision and as the basis for evaluation Dice scores.","marker":"Wasserthal et al., 2023"},{"why":"The recent foundation-model baseline compared throughout; its reported training-data imbalance explains part of the regional performance differences.","marker":"Tian et al., 2024"},{"why":"Elastix, the generic classical iterative registration baseline the model is measured against.","marker":"Klein et al., 2010"},{"why":"VoxelMorph-style unsupervised loss scheme (image similarity plus deformation regularization) that underlies the training procedure.","marker":"Balakrishnan et al., 2019"},{"why":"Standard UNet architecture that all registration blocks reuse, supporting the lightweight and easy-to-reproduce claim.","marker":"Ronneberger et al., 2015"},{"why":"Source of the private multiphase liver dataset and an organ-specific cascaded registration method used as an external comparison point.","marker":"Pham et al., 2024"},{"why":"Provides the public thoracic and abdominal evaluation datasets used to test out-of-distribution generalizability.","marker":"Hering et al., 2023"},{"why":"Recursive cascaded registration idea that the field decomposition strategy extends to region-specific whole-body blocks.","marker":"Zhao et al., 2019"}],"fun_headline_variants":["Whole-body CT registration rivals single-organ models","TotalRegistrator: one model, many organs","Lightweight CT registration matches specialist models","UNet-based whole-body CT registrator beats baselines","General CT registration from 591 scan pairs"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The model's training depends on organ segmentation masks that were generated automatically and never manually corrected; if those masks carry systematic errors, the learned deformation fields and the reported Dice improvements are biased.","fun_headline_variants_meta":{"raw":{"variants":["Whole-body CT registration rivals single-organ models","TotalRegistrator: one model, many organs","Lightweight CT registration matches specialist models","UNet-based whole-body CT registrator beats baselines","General CT registration from 591 scan pairs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000112,"raw_usage":{"total_tokens":906,"prompt_tokens":763,"completion_tokens":143,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":87}},"tokens_in":507,"tokens_out":143,"duration_ms":2794,"temperature":1.0,"reasoning_tokens":87,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:58:02.045417+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Manually correct the training-time organ masks for a subset of cases, retrain under identical settings, and compare held-out Dice; if scores change materially, the automatic labels were load-bearing. Alternatively, test on whole-body CT scans from scanners and patient populations not represented in the training cohort; a large drop in abdominal Dice would falsify the generalizability claim.","supporting_citations":[],"review_version":1}