{"id":"3c431926-bc16-45ac-abea-194294acb697","arxiv_id":"2411.13873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Sli2Vol+ learns slice-to-slice correspondences with guidance from pseudo-labels, improving single-slice-annotated 3D segmentation over the Sli2Vol baseline by about 5.6 Dice points on CT and 3.5 on MRI.","lead":"This paper presents Sli2Vol+, a method that segments 3D medical images from just one labeled slice per volume by learning to propagate labels across neighboring slices. It could reduce the high annotation cost of medical imaging while generalizing across organs, CT/MRI modalities, and MRI pulse sequences.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cross-modal MRI experiment (Table 3) does not report a patient-level train/test split, so if the same Decathlon patients appear in T1w training and FLAIR/T1gd/T2w testing, the reported cross-modal gains may reflect patient overlap instead of learned correspondence generalizability.","rationale":"I do not find the pseudo-label-quality assumption to be the most fragile point. The CT experiments in Table 2 and the ablations in rows (g)-(i) give empirical support that adding PLs to the OEG-CFN improves propagation over Sli2Vol, and the manuscript never claims the PLs are close to ground truth; it claims only that they provide useful object estimation guidance. Even if PL quality is not directly measured, the consistent gains suggest the mechanism is working. The most load-bearing weakness is in the cross-modal MRI evaluation, which is a central part of the claimed generalizability. Section 4.1 and Table 3 describe training on T1w and testing on FLAIR, T1gd, and T2w of the Decath-Brain Tumours dataset, but without a stated patient-level split, the multi-modal same-patient nature of that dataset can confound the results. If the same patients appear in training and test, the model has prior access to the exact anatomy being segmented, making the cross-modal numbers uninformative about generalization to unseen patients. This is a concrete, checkable threat to a headline claim, unlike the slice-selection protocol or missing hyperparameters, which are secondary and addressable. I therefore keep a conditional verdict but add a specific condition: demonstrate patient-disjoint splitting and rerun the cross-modal experiment. If the split is already disjoint, the concern is resolved and the verdict remains as originally assessed.","tokens_in":16331,"tokens_out":8017,"duration_ms":88547,"concrete_test":"Extract the patient identifiers for the Decath-Brain Tumours volumes used in the T1w training set and the FLAIR/T1gd/T2w test sets. If any patient appears in both, rerun Table 3 with a strict patient-disjoint split (e.g., train on T1w volumes from one set of patients, test on FLAIR/T1gd/T2w volumes from held-out patients), and recompute the mean Dice for row (i) versus rows (e) and (f). If the margin over Sli2Vol and Vol2Flow shrinks to within one standard deviation, the cross-modal generalizability claim is not supported by the current protocol.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline claim includes generalizability across modalities and modals, and Table 3 is the main evidence for cross-modal performance: Sli2Vol+ reaches 54.1 mean Dice versus 50.6 for both Sli2Vol and Vol2Flow. However, Section 4.1 states only that the model is trained on the Decath-Brain Tumours T1w modal and tested on the FLAIR, T1gd, and T2w modals, without specifying any patient-level split. Since the Decathlon Brain Tumours dataset consists of multi-modal acquisitions of the same patients, a volume-level split without patient exclusion would put the same anatomy in both the T1w training set and the cross-modal test sets. The model could then exploit patient-specific shape and location information during test-time propagation, inflating the reported cross-modal Dice and the claimed advantage over prior methods. This is not addressed by any ablation or stated data-splitting protocol, and it directly affects the abstract's 'different modals' claim and the conclusion's generalizability statement. The concern is conditional: if the authors already used a strict patient-disjoint split, it does not land, but the manuscript currently gives no basis for that assumption.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Sli2Vol+, a self-supervised framework for 3D medical image segmentation from a single annotated slice per training and testing volume. The method first generates pseudo-labels (PLs) for training volumes using Sli2Vol, refines them with UNETR++, and then trains an Object Estimation Guided Correspondence Flow Network (OEG-CFN) that learns slice-to-slice correspondences from both consecutive slices and PLs. A gradient-enhanced image generator (GEIG) is introduced as an information bottleneck for the slice reconstruction pretext task. At test time, the trained network propagates the single annotated slice to the rest of the volume. The authors report experiments on nine public datasets (CT and MRI) covering ten structures, with ablations showing incremental Dice improvements from each component, and comparisons against fully supervised, single-slice supervised, and mask propagation baselines.","tokens_in":16625,"tokens_out":6369,"duration_ms":62285,"significance":"If the results hold, the paper addresses a practically important problem: reducing the annotation burden for 3D medical segmentation. The evaluation is extensive (nine datasets, multiple organs, CT/MRI, cross-modal MRI), includes five-seed repetitions with standard deviations, and the code is publicly released. The ablation study is carefully structured and each proposed component (PLs, refinement, GEIG) yields a consistent improvement. The additional robustness experiment using ScribblePrompt-generated annotations (Table 5) is a strength. However, the central claim hinges on the reliability of the pseudo-labels and on the cross-modal MRI experiment, both of which currently lack crucial validation or specification.","major_comments":[{"comment":"The cross-modal MRI experiment does not state whether the T1w training volumes and the FLAIR/T1gd/T2w test volumes come from disjoint patients. The Decathlon Brain Tumours dataset contains multi-modal acquisitions of the same patients, so a volume-level split without patient exclusion would place the same anatomy in both training and test sets, inflating the reported cross-modal Dice scores. The authors must specify the patient-level split; if it is not patient-disjoint, the experiment must be re-run with a strict patient-exclusion protocol.","section":"§4.1, Table 3"},{"comment":"The inference procedure is under-specified. During training, OEG-CFN computes the affinity matrix from features of both the slices and the PLs, but at test time PLs are unavailable. The paper does not explain how the trained two-path network is adapted for test-time affinity computation (e.g., zero-filling the PL inputs, using only the slice path, or training a single-path variant). Without this information, the reported test-stage results cannot be reproduced or fully assessed.","section":"§3.4, Fig. 1"},{"comment":"The gradient-enhanced image generator depends on unspecified hyperparameters: d (number of directions) and s (number of scales), and the exact operational definitions of 'direction' and 'scale' are not provided. These values are not reported in §4.2 or elsewhere. The authors should give the concrete settings used in all experiments; otherwise the claimed benefit of GEIG cannot be independently evaluated.","section":"§3.3, Eq. (3)"},{"comment":"The quality of the pseudo-labels, which are the key supervisory signal of the framework, is never quantified. No Dice score or other metric is reported for the generated or refined PLs against ground truth on the training volumes (or a held-out subset). Since the method's improvement is attributed to 'object estimation guided' learning, the authors should provide PL quality statistics (e.g., mean Dice before and after UNETR++ refinement) to validate the underlying mechanism.","section":"§3.2"}],"minor_comments":[{"comment":"The term 'modals' is used repeatedly; consider using 'modalities' or explicitly defining 'modals' as MRI sequences (e.g., T1w, T2w, FLAIR).","section":"Throughout"},{"comment":"The claim that Sli2Vol+ and FS-SS use 'the same amount of annotations' is inaccurate because Sli2Vol+ additionally uses one annotated slice per training volume; revise the wording to reflect the actual annotation budgets.","section":"§5.1, point (2)"},{"comment":"Add the number of training volumes for C4KC-KiTS, CT-LN, and CT-Pancreas to provide a complete picture of the annotation cost.","section":"Table 2"},{"comment":"The selection rule for the annotated slice is described only for test volumes ('±3 slices around the slice with the largest GT annotation'); specify how the single annotated slice per training volume is chosen.","section":"§4.3"},{"comment":"The conclusion claims the method segments 'any anatomical structures,' but experiments cover only ten structures; soften the generality claim.","section":"§6"},{"comment":"The statement 'batch size ... maximum size allowed by the GPU' is not reproducible; give a concrete value or a precise heuristic used to determine the batch size.","section":"§4.2"},{"comment":"Define the notation 'Sli2Vol+ (using ScribblePrompt)' in the caption or text, clarifying that only the training-volume slice annotations are produced by ScribblePrompt while test annotations remain manual.","section":"Table 5"},{"comment":"The label 'GTV' in the top row is not defined in the text; explain its meaning or replace it with a clearer term.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The most critical issue is the patient-level split in the cross-modal MRI experiment (Table 3); if the split is already patient-disjoint, the paper could be acceptable after clarifying the inference protocol and implementation details. I would also encourage the authors to consider an annotation-matched comparison with Sli2Vol (e.g., providing Sli2Vol with the same additional training annotations) to better isolate the contribution of the proposed OEG-CFN and GEIG components."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does honest incremental work in the single-slice-annotation setting. The new pieces are OEG-CFN, which concatenates slice features with pseudo-label features to learn correspondences, and GEIG, which replaces Sli2Vol's first-order edge profile with softmax-normalized multi-scale second-order derivatives. The ablations show each added component nudges Dice upward, and the evaluation across nine datasets with five seeds and t-tests is more careful than most work in this area. The gains over Sli2Vol (about 5.6 Dice on CT, 3.5 on MRI) are modest but real. Using UNETR++ to refine pseudo-labels is a sensible self-training step and the paper gives it credit.\n\nThe soft spot that matters most is the cross-modal MRI experiment in Table 3. The model is trained on the T1w modal and tested on FLAIR, T1gd, and T2w modals of the same Decathlon brain tumour dataset, but the paper never states whether the split is patient-disjoint. Since that dataset contains multi-modal acquisitions of the same patients, a volume-level split could put the same anatomy in both training and test. That would inflate the reported Dice and undercut the 'different modals' claim. This is a conditional concern; if the authors already split by patient, it does not land. But the manuscript needs to say so explicitly.\n\nTwo smaller issues. First, pseudo-label quality is never quantified, so we are trusting that the refined PLs are accurate enough to teach useful correspondences. Second, the slice selection protocol (always within +-3 slices of the largest GT cross-section) is inherited from Sli2Vol and favors favorable starting points; it should be stated as a limitation. The missing GEIG hyperparameters (d and s) and ConvNet architecture details are minor and easily fixed. The 'any anatomical structures' phrase is overbroad for ten SOIs, but that is standard abstract inflation.\n\nThe central argument, that PL-guided correspondence learning addresses error drift and discontinuity, holds up based on the ablations. The cross-modal split is the one load-bearing question mark. This paper deserves peer review, not desk rejection, provided the authors can confirm the patient-disjoint split and supply the missing implementation details. I would send it out with a request for that confirmation.","headline":"Solid incremental extension of Sli2Vol with real gains on sparse-annotation 3D segmentation; the cross-modal MRI result needs a patient-disjoint split before the generalizability claim holds.","tokens_in":738,"tokens_out":748,"would_cite":true,"duration_ms":23862,"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":"Sli2Vol+ segments whole 3D medical volumes from a single labeled slice.","keywords":["self-supervised learning","medical image segmentation","mask propagation","pseudo-labels","slice correspondence","3D segmentation","single-slice annotation","affinity matrix"],"falsifier":"Replace the refined pseudo-labels used to train OEG-CFN with random binary masks (preserving object-slice statistics) and rerun the CT experiments: if Dice does not drop substantially, the object-estimation guidance is not the source of the improvement.","tokens_in":16168,"feed_emoji":"🩻","tokens_out":6951,"duration_ms":58697,"temperature":0.7,"pith_summary":"Sli2Vol+ is a self-supervised framework for segmenting 3D medical images using only a single annotated 2D slice per training and test volume. The paper argues that previous mask-propagation methods, which learn correspondences between neighboring slices by slice reconstruction, accumulate errors as they propagate and fail when anatomical structures appear or disappear between slices. To fix this, Sli2Vol+ first generates and refines pseudo-labels for training volumes, then trains a network to learn correspondences from both consecutive slices and these pseudo-labels, so the object estimates guide the correspondence learning. The authors report that this reduces error drift and handles discontinuities, outperforming existing propagation methods across nine public CT and MRI datasets and ten structures of interest.","feed_headline":"Sli2Vol+ segments whole 3D volumes from one labeled slice","feed_subtitle":"Pseudo-label-guided correspondence learning cuts error drift and beats prior mask propagation on nine CT/MRI datasets.","key_machinery":"The load-bearing component is the Object Estimation Guided Correspondence Flow Network (OEG-CFN), which learns an affinity matrix between a slice and its neighbor from two parallel feature streams: one computed from the two consecutive slices and one from their corresponding pseudo-labels, with the two ConvNets sharing weights. The affinity matrix is computed as in Sli2Vol, a local softmax over dot products between query and key features, and is used at test time to propagate a segmentation mask slice by slice. The pseudo-labels themselves are produced by first running Sli2Vol from the single annotated slice and then refining the resulting masks with a 3D segmentation network (UNETR++), so that the object estimates incorporate global 3D context. A second piece of machinery is the gradient-enhanced image generator (GEIG), which replaces Sli2Vol's edge-profile generator: it encodes each pixel by a softmax-normalized histogram of second-order derivatives at multiple directions and scales, making the reconstruction pre-text task less noise-sensitive and better at localizing edges.","core_discovery":"The central claim is that injecting object-level supervision in the form of pseudo-labels into the slice-correspondence learning process yields more reliable correspondences than learning purely from slice reconstruction, and that these correspondences transfer to test volumes where no pseudo-labels are available. The paper demonstrates that an Object Estimation Guided Correspondence Flow Network (OEG-CFN), which concatenates key and query features from slices with those from pseudo-labels before computing an affinity matrix, improves mask propagation over the Sli2Vol baseline by an average of 5.6 Dice points on CT and MRI datasets. The paper further claims that refining pseudo-labels with a 3D network and replacing Sli2Vol's edge-profile input with gradient-enhanced images (histograms of second-order derivatives) each add further gains, and that the full pipeline generalizes across organs, modalities, and datasets with only about a 7-point drop versus fully supervised same-domain models under domain shift.","pith_inferences":["If the reported gains are driven by pseudo-label quality, then an experiment that corrupts the refined pseudo-labels (for example, by erasing random object regions) should degrade final segmentation in proportion to the corruption, a dependence the paper does not test directly.","The framework's reliance on a single annotated slice per training volume makes it a natural bootstrapping engine for new volumes: the same OEG-CFN could propagate a few human-corrected slices interactively, an application the paper mentions only in passing.","Because GEIG decouples the representation from raw intensities, the method may transfer more readily between MRI sequences or CT protocols than intensity-based reconstruction methods; a dedicated cross-protocol experiment with paired scans would settle this."],"forward_implications":["With one annotated slice per training and test volume, Sli2Vol+ reports mean Dice of 78.9 across the CT organ suite and 54.1 across cross-modal brain-tumor MRI, exceeding Sli2Vol, Vol2Flow, and VoxelMorph on every dataset tested.","The pseudo-label guidance specifically reduces error accumulation: qualitative results show improved propagation at slices far from the labeled slice and at slices where an organ ends.","The annotation cost is low enough that the single training-volume slice can be generated automatically by an interactive or foundation segmentation tool (e.g., ScribblePrompt) with little loss in performance.","Across cross-domain and cross-modality evaluations, the method keeps performance within roughly 7 Dice points of same-domain fully supervised models, suggesting the learned correspondences generalize rather than memorize training volumes."],"supporting_citations":[{"why":"Supplies the base slice-reconstruction method and generates the initial pseudo-labels.","marker":"[59]"},{"why":"Refines the pseudo-labels using global 3D context.","marker":"[43]"},{"why":"VoxelMorph is a primary registration-based baseline.","marker":"[2]"},{"why":"Vol2Flow is a primary registration-flow baseline.","marker":"[3]"},{"why":"Provides the Decathlon test datasets for CT and MRI evaluation.","marker":"[45]"},{"why":"ScribblePrompt generates the single-slice annotations used to show annotation cost can be automated.","marker":"[53]"}],"fun_headline_variants":["Segment 3D volumes with a single labeled slice","Object-guided correspondence flow tames slice propagation drift","Pseudo-labels power one-slice 3D segmentation","Sli2Vol+ cuts error accumulation in mask propagation","From one slice to full 3D: Sli2Vol+ does it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The refined pseudo-labels are accurate enough that training the correspondence network to imitate them teaches transferable correspondences, even though no pseudo-labels are available at test time.","fun_headline_variants_meta":{"raw":{"variants":["Segment 3D volumes with a single labeled slice","Object-guided correspondence flow tames slice propagation drift","Pseudo-labels power one-slice 3D segmentation","Sli2Vol+ cuts error accumulation in mask propagation","From one slice to full 3D: Sli2Vol+ does it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":1985,"prompt_tokens":1081,"completion_tokens":904,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":820}},"tokens_in":697,"tokens_out":904,"duration_ms":8124,"temperature":1.0,"reasoning_tokens":820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:03.762939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the refined pseudo-labels used to train OEG-CFN with random binary masks (preserving object-slice statistics) and rerun the CT experiments: if Dice does not drop substantially, the object-estimation guidance is not the source of the improvement.","supporting_citations":[{"cited_title":"Sli2V ol: Annotate a 3D volume from a single slice with self- supervised learning","cited_arxiv_id":null,"evidence_quote":"Supplies the base slice-reconstruction method and generates the initial pseudo-labels."},{"cited_title":"V oxelMorph: A learning frame- work for deformable medical image registration","cited_arxiv_id":null,"evidence_quote":"VoxelMorph is a primary registration-based baseline."},{"cited_title":"V ol2Flow: Segment 3D volumes using a sequence of registration flows","cited_arxiv_id":null,"evidence_quote":"Vol2Flow is a primary registration-flow baseline."}],"review_version":1}