{"id":"d135e725-fb3a-4805-bc65-ec1cfbe42106","arxiv_id":"1908.08870","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A topology-preserving augmentation pipeline that uses a cardiac contiguous topology template reduces topological errors in CNN-based segmentation of congenital heart defects in 3D CMR.","lead":"The paper presents a data augmentation pipeline that keeps the connectivity of heart tissue labels intact when training images are warped for CNN segmentation. It shows the pipeline reduces spurious heart defects in segmentations, which matters for creating reliable 3D printed models for surgery.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Leakage risk: the 10,000 augmented examples were pre-computed from all ten cases with no stated per-fold regeneration; if test-subject augmentations enter training, the reported topological improvement may be an artifact.","rationale":"The reader's formal 'weakest_assumption' field names the CCT template assumption, but the reader's own rationale identifies the augmentation leakage risk as the most serious issue. I agree with that assessment, and the full text supports it: the augmentation examples are pre-computed from all ten cases with no per-fold regeneration statement. The central numerical claim therefore rests on an unverified protocol detail. This is a concrete, testable threat to the paper's central claim, and the conditional verdict already given by the reader is the appropriate one. I do not change the verdict because the reader already required clarification of this point; my stress-test confirms that the concern is load-bearing and should be resolved before the quantitative claim is accepted. The strongest independent support in the paper is the small effect size and the reproducibility-friendly use of a public dataset, but without code or a per-fold description the leakage concern cannot be dismissed. The CCT template assumption remains a secondary concern, but it is secondary because even a clinically perfect template would not validate a comparison made on contaminated training data.","tokens_in":5805,"tokens_out":3075,"duration_ms":34701,"concrete_test":"none","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that topology-preserving augmentation, combined with clinically justified transformations, significantly reduces topological errors in CNN-based CHD segmentation. The evidence is the comparison in Figure 4: false-positive topological clusters fall from 12.0 (6.3,22.5) with naive clinically justified augmentation to 6.5 (6.0,10.5) with the topology-preserving pipeline, with Wilcoxon p = 0.022. That result is only valid if training and test subjects are strictly separated at the level of augmentation examples. The paper describes five-fold cross-validation with train-on-eight, test-on-two, but the Implementation section states: 'From the ten cases provided by the HVSMR Challenge, a total of 10,000 training examples were pre-computed by data augmentation according to figure 2.' No statement indicates that augmentation was regenerated per fold using only the eight training subjects. If the pre-computed pool was built from all ten subjects and then split by subject into folds, the augmented replicas of the two test subjects may still appear in the training pool, since random crops are drawn from the augmented data during training. Because the topology-preserving pipeline corrects label maps to preserve the specific CCT of each subject, a test subject's transformed and corrected labels in the training pool would directly expose the network to the test topology it is later asked to predict. This leakage would contaminate both absolute performance and the comparison between augmentation schemes, since corrected test-subject labels would be easier to memorize than the topologically corrupted naive labels. The paper's own reported change of less than 0.5% of blood pool voxels makes the correction subtle enough that leakage could plausibly drive the measured effect. This is the load-bearing weakness because no subsequent analysis of the metric or the CCT template can rescue a contaminated training/test split.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript addresses the problem of topological errors in CNN-based segmentation of congenital heart defects from 3D CMR, motivated by patient-specific 3D printing. It proposes a topology-preserving augmentation pipeline that applies clinically justified spatial transformations to the image and label maps, then uses a fast-marching algorithm to enforce the 'cardiac contiguous topology' (CCT) of the ground truth labels on the transformed label maps. The authors define CCT as the connectivity between cardiac sub-classes after each is made ball-like, and construct a one-voxel skeleton template via erosion. They also introduce a novel metric that counts clusters of topologically relevant disagreement. Using five-fold cross-validation on the ten HVSMR training cases, they compare clinically justified augmentation with and without topology preservation, and report that the topology-preserving version reduces the median number of false-positive topological clusters from 12.0 to 6.5 (Wilcoxon p=0.022) and improves DSC for blood pool and myocardium.","tokens_in":6093,"tokens_out":4863,"duration_ms":48945,"significance":"If the central claim holds, this is a valuable contribution: it identifies a realistic failure mode of conventional augmentation (topological corruption of thin tissue interfaces), provides a principled correction, and introduces a metric more sensitive than DSC to clinically relevant interface errors. The CCT notion is an elegant way to handle arbitrarily complex combinations of defects, and the authors make a plausible argument for why previous augmentation choices (orthogonal rotations) were suboptimal. However, the evidence base is small (ten subjects, two per test fold), the metric is not externally validated, and the leakage ambiguity described below must be resolved before the empirical claim can be accepted.","major_comments":[{"comment":"The Implementation section states that 'From the ten cases provided by the HVSMR Challenge, a total of 10,000 training examples were pre-computed by data augmentation according to figure 2.' It does not state whether the augmentation was regenerated for each of the five folds using only the eight training subjects of that fold. If the pre-computed pool contains augmented examples of a test subject, then during that fold's training the network can see transformed and topology-corrected versions of the subject it is later tested on, directly exposing it to the test subject's CCT. This would invalidate the comparison in Figure 4 and the Wilcoxon test (p=0.022). The authors must either confirm that per-fold separation was enforced at the augmentation level, or re-run the experiments with the augmentation generated separately for each training fold.","section":"Section 2.2, Implementation"},{"comment":"The pipeline enforces the topology of a CCT template obtained by (i) correcting each cardiac sub-class to ball topology and (ii) eroding a well-composed blood pool to a one-voxel skeleton. This assumes that the simplified CCT retains all clinically relevant connections and that no important connection is removed or distorted by the erosion. If, for example, a thin but real communication is thinner than one voxel after resampling, the template could break that connection and the augmentation would then train the network to incorrectly separate connected chambers. The manuscript should provide evidence (e.g., expert review of the templates, or quantitative comparison of CCT with the original labels) that the CCT construction preserves the clinically relevant topology in all ten cases.","section":"Section 2.1, CCT template construction"},{"comment":"The proposed metric counts 'topologically relevant clusters of voxels where inferred and ground truth segmentations disagree' and uses this to measure the 'anomalous appearance of congenital defects.' No validation is provided that these clusters correspond to clinically meaningful defects, nor is the metric compared with an expert assessment or a more established topological descriptor. Since the paper's main quantitative finding is expressed in this metric, the manuscript should include at least an illustrative case-by-case breakdown and ideally a small expert study to support the clinical relevance of the metric.","section":"Section 2.2, Metrics"}],"minor_comments":[{"comment":"The figure caption does not define the abbreviations 'CJ' and 'TP' in printed form; the caption should explicitly state that CJ stands for clinically justified, TP for topology-preserving, and *CJ + TP for the proposed pipeline.","section":"Figure 4"},{"comment":"The term 'well-composed set' is introduced without a citation or formal definition; please provide a reference or a clear definition.","section":"Section 2.1"},{"comment":"The description of the fast-marching topology correction is brief; please specify the parameters used (e.g., neighborhood connectivity and stopping criterion) so that the experiments are reproducible.","section":"Section 2.2, Implementation"},{"comment":"The DSC improvements are reported without confidence intervals or statistical tests; please report these to support the claim of benefit.","section":"Results and discussion"},{"comment":"Minor typo: 'MIPA V' should likely be 'MIPAV'.","section":"Section 2.2, Implementation"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is an interesting idea with a plausible mechanism, but the empirical evaluation is currently too fragile to support the central claim. The lack of clarity on whether augmentation examples were separated per fold is the most serious issue; if the authors cannot confirm leakage-free training, the main result may be an artifact. I recommend major revision, with the expectation that the authors either provide the missing clarification and re-run the experiments, or add validation of the CCT assumption and the proposed metric."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper introduces a genuinely nice idea—using fast-marching topology correction to keep augmented label maps consistent with the original cardiac topology—and supports it with a new, clinically interpretable cluster-counting metric. The writing is clear, the figures are helpful, and the clinical motivation (thin septal interfaces and 3D printing) is real.\n\nThe problem is the training protocol. The authors say 10,000 training examples were \"pre-computed by data augmentation\" from the ten HVSMR cases. They then run five-fold cross-validation (train on eight, test on two) and randomly crop batches from the augmented data. No statement says that the augmented pool was regenerated per fold from only the eight training subjects. If the pool was built once from all ten cases, then for each fold the two test subjects' augmented replicas—including their topology-corrected labels—are sitting in the training pool. The network would then be directly exposed to the test topology it is later asked to predict. That would explain the reported drop in false-positive clusters (12.0 to 6.5, p=0.022) without any real learning. This is a load-bearing flaw, not a nuisance.\n\nEverything else is softer. The dataset is tiny (ten cases), so confidence intervals are wide; the DSC gains are modest (e.g., +0.007 for blood pool); and the new metric isn't externally validated. No code or trained models are provided, so reproduction would require reimplementation. None of these alone would sink the paper, but the leakage ambiguity is enough to withhold acceptance.\n\nI hope the authors can show that augmentation was done per fold. If they can, the paper is a reasonable contribution: it's the first to my knowledge to combine topology correction with augmentation for CHD segmentation, and the CCT template idea is worth building on. If they can't, the empirical comparison is meaningless.\n\nVerdict: deserves peer review, but as a conditional—authors must clarify the augmentation split and ideally make the augmented data available. I'd bring it to a reading group to discuss the design, not because the result is trustworthy yet.","headline":"Nice idea and a real clinical motivation, but the training protocol has a plausible data-leakage ambiguity that could invalidate the headline result.","tokens_in":6662,"tokens_out":3080,"would_cite":false,"duration_ms":28124,"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":"This paper shows that enforcing ground-truth label topology during data augmentation—via fast-marching correction against a cardiac-topology template—cuts false defect clusters in congenital heart segmentation from 12.0 to 6.5 (p=0.022).","keywords":["image segmentation","data augmentation","topology preservation","congenital heart disease","cardiac magnetic resonance","fast marching","convolutional neural network","3D printing"],"falsifier":"Take one labelled volume with a known thin connection between two chambers, such as an atrial septal defect; construct its CCT template and check whether the one-voxel skeleton still contains that connection. If the erosion step severs the channel, then enforcing the template topology during augmentation would actively train the network to delete the very defect the segmentation is meant to show.","tokens_in":5581,"feed_emoji":"🫀","tokens_out":11003,"duration_ms":98813,"temperature":0.7,"pith_summary":"This paper identifies a failure mode in CNN training for congenital heart disease segmentation: standard data augmentation transforms and resamples ground-truth label maps, and nearest-neighbour resampling can create or destroy thin tissue interfaces, producing spurious defects in the training labels. The authors propose a pipeline that builds a simplified topological skeleton of each label map, the cardiac contiguous topology (CCT) template, transforms it alongside the labels, and uses fast-marching topology correction to force the resampled blood pool back to the template's topology. In five-fold cross-validation on ten paediatric cardiac MR cases, coupling clinically realistic augmentation with this correction lowers the median number of topologically relevant false-positive error clusters from 12.0 to 6.5, a statistically significant improvement (p=0.022). If correct, the work establishes that label-map topology is a trainable property and that clinically meaningful augmentation does not have to corrupt it.","feed_headline":"Topology-safe augmentation cuts heart-defect errors from 12 to 6.5","feed_subtitle":"Fast-marching fix during augmentation cuts false defect clusters from 12 to 6.5.","key_machinery":"The central mechanism is the CCT template: a one-voxel-wide skeleton of the ground-truth blood pool that records how the heart's chambers and vessels are connected, with each sub-class first simplified to ball topology so that trabeculation is removed. The companion mechanism is a fast-marching topology-correction algorithm, which flips only non-simple voxels—voxels whose label can be changed without changing the topology of the object or its complement—until the resampled blood pool has the same topology as the transformed CCT template. Together they let the pipeline repair whatever topological damage the resampling step caused, while changing almost no voxels.","core_discovery":"When ground-truth labels for the cardiac blood pool are spatially transformed and resampled by nearest-neighbour interpolation during data augmentation, thin interfaces such as the atrial septum can develop spurious holes or connections, which appear in the final segmentation as anomalous congenital defects. The authors' pipeline first constructs, for each training label map, a cardiac contiguous topology (CCT) template: each cardiac sub-class is corrected to ball topology and the recombined blood pool is topologically eroded to a one-voxel skeleton, so the template records only clinically meaningful chamber-to-chamber contiguity. After the image transform, the blood pool label is resampled with trilinear interpolation and corrected with a fast-marching algorithm until its topology matches the transformed CCT template. In five-fold cross-validation on ten paediatric CMR cases, CNNs trained with clinically justified transformations plus this correction produced a median of 6.5 topologically relevant false-positive error clusters, down from 12.0 for the same transformations without topology correction (Wilcoxon p=0.022), and fewer than 0.5% of blood-pool voxels were altered.","pith_inferences":["The same template-plus-correction construction should transfer to any segmentation task whose foreground can be expressed as several well-composed contiguous sub-classes, such as vascular trees, airways, or multi-chamber organs; the only requirement is a clinical definition of which connections count.","A direct next step is to use the paper's topological error metric as a loss term or early-stopping signal during training; the reported results suggest that penalising topologically relevant error clusters could suppress the remaining ~6.5 false-positive clusters further.","The authors are explicit that the myocardium label is still resampled by nearest-neighbour interpolation because its true topology is only partially visible in CMR; extending the topology guarantee to the myocardium would require a clinically agreed definition of muscle topology and would change the correction statistics.","Because the pipeline changes less than 0.5% of blood-pool voxels, the corrective step is acting almost entirely on the handful of voxels that determine thin-interface topology; this suggests a topology-preserving resampler could eventually be implemented as an end-to-end differentiable layer rather than a post-hoc correction."],"forward_implications":["Clinically realistic augmentation and label-topology preservation become compatible rather than conflicting objectives.","Median topologically relevant false-positive clusters fall from 12.0 to 6.5 when the correction is applied (Wilcoxon p=0.022), beating both naive clinically justified augmentation and the orthogonal-rotation-only augmentation used in previous work.","The correction is cheap in voxel terms: fewer than 0.5% of blood-pool voxels are altered, and blood-pool overlap is preserved (DSC 0.925 vs 0.918).","Topological correctness can be evaluated separately from overlap, giving a clinically interpretable count of where a segmentation would falsely suggest a defect."],"supporting_citations":[{"why":"Supplies the fast-marching topology-correction algorithm that removes non-simple points to enforce the template topology on resampled labels.","marker":"[1]"},{"why":"Defines the V-net 3D CNN architecture used for all segmentation experiments.","marker":"[4]"},{"why":"Provides the ten paediatric CMR cases with manual whole-heart labels and acquisition details that the study and its expert corrections build on.","marker":"[5]"},{"why":"Prior CNN segmentation of congenital heart disease that restricted augmentation to topology-preserving orthogonal rotations; a key baseline the paper improves upon.","marker":"[6]"},{"why":"Prior densely-connected volumetric CNN that also limited augmentation to orthogonal rotations and lateral inversion; supplies the comparison for clinically unjustified augmentation.","marker":"[7]"},{"why":"Prior 3D FractalNet work with the same augmentation restriction; another baseline for both overlap and topological performance.","marker":"[8]"},{"why":"Prior deeply-supervised 3D FCN whole-heart segmentation result used when situating the reported overlap scores.","marker":"[2]"}],"fun_headline_variants":["Heart defect errors halved by topology-safe augmentation","Fast-marching fix cuts heart-defect errors from 12 to 6.5","Topology-correct augmentation slashes false defects in CHD","Topology-aware augmentation halves false heart defects","Topology-preserving augmentation cuts CNN heart errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the CCT template—obtained by correcting every cardiac sub-class to ball topology and eroding the blood pool to a one-voxel skeleton—preserves exactly the clinically meaningful connections, so that any connection missing from the template is a defect rather than a real anatomical feature.","fun_headline_variants_meta":{"raw":{"variants":["Heart defect errors halved by topology-safe augmentation","Fast-marching fix cuts heart-defect errors from 12 to 6.5","Topology-correct augmentation slashes false defects in CHD","Topology-aware augmentation halves false heart defects","Topology-preserving augmentation cuts CNN heart errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3287,"prompt_tokens":986,"completion_tokens":2301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2220}},"tokens_in":602,"tokens_out":2301,"duration_ms":15276,"temperature":1.0,"reasoning_tokens":2220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:26:46.582748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one labelled volume with a known thin connection between two chambers, such as an atrial septal defect; construct its CCT template and check whether the one-voxel skeleton still contains that connection. If the erosion step severs the channel, then enforcing the template topology during augmentation would actively train the network to delete the very defect the segmentation is meant to show.","supporting_citations":[{"cited_title":"Computer methods and programs in biomedicine 88(2), 182--190 (2007)","cited_arxiv_id":null,"evidence_quote":"Supplies the fast-marching topology-correction algorithm that removes non-simple points to enforce the template topology on resampled labels."},{"cited_title":"In: 3D Vision (3DV), 2016 Fourth International Conference on","cited_arxiv_id":null,"evidence_quote":"Defines the V-net 3D CNN architecture used for all segmentation experiments."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Prior densely-connected volumetric CNN that also limited augmentation to orthogonal rotations and lateral inversion; supplies the comparison for clinically unjustified augmentation."}],"review_version":1}