{"id":"bf50bdee-32e8-4164-ba0b-5109c5ed931d","arxiv_id":"2412.15526","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SGTC uses CLIP semantic guidance and three-view co-training to segment medical volumes from just three annotated slices per volume, with gains over baselines mainly driven by the extra label.","lead":"This paper presents SGTC, a semi-supervised method that segments 3D medical volumes using only three hand-annotated slices per training volume, combining CLIP text guidance with co-training among three networks. It reports large gains over prior sparse-annotation methods on three public datasets, but the headline comparisons do not control for the extra annotated slice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed SOTA edge may be driven by an extra annotation slice; matched-budget Table 5 shows only +0.2 Dice over BCP, with no significance test.","rationale":"The reader's stated weakest assumption is the annotation protocol's reliance on informed slice selection, which is acknowledged in the paper's limitations. However, the reader's rationale also emphasizes the unequal annotation budget and the minimal advantage in the controlled comparison—this is the more load-bearing concern for the central claim. I partially agree: the slice-selection dependence is a real practical limitation, but the primary threat to the claim of state-of-the-art performance is that the main tables compare three-slice supervision against two-slice supervision, and the only matched-budget experiment shows a negligible, untested difference. The LiTS unlabeled-count discrepancy further undermines confidence. My proposed test directly isolates the annotation-budget effect and would determine whether the method's edge is intrinsic or merely a consequence of more labels. I do not see internal inconsistency in the method itself; the concern is about experimental support. Therefore the correct verdict remains CONDITIONAL, pending matched-budget, statistically valid comparisons. I set verdict_should_be to UNCHANGED because the reader already recommended CONDITIONAL, and my analysis reinforces that recommendation without moving it further.","tokens_in":13252,"tokens_out":3298,"duration_ms":31102,"concrete_test":"On KiTS19 with 19 labeled volumes, train BCP, Desco, and SGTC using the exact same three orthogonal slices per volume (i.e., provide the baselines the extra sagittal slice), with 3 random seeds each; report Dice, HD, and ASD with paired Wilcoxon tests. If SGTC's advantage over BCP is not consistently significant (p<0.05), the headline SOTA claim should be downgraded to 'competitive under matched budget'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that SGTC outperforms state-of-the-art semi-supervised methods under sparse annotation—rests on Tables 1-3, where SGTC uses three orthogonal slices (SCA) while every baseline uses only two slices (CA). This unequal annotation budget conflates method quality with extra supervision. Table 5 is the only controlled comparison: with all methods given the same CAC or SCA strategy on KiTS19 (9 labeled volumes), SGTC reaches 0.915 Dice vs BCP's 0.913, a 0.2-point gap with no error bars or significance test. Additionally, Table 3 lists SGTC(Dual) with 80 unlabeled scans while all other rows use 90, an unexplained inconsistency that weakens the LiTS comparison. The paper's own limitation note admits performance degrades when selected slices contain limited foreground, so the protocol requires informed slice selection. Together these issues mean the reported superiority is not yet robustly established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SGTC, a semi-supervised framework for medical image segmentation under a sparse annotation protocol in which each of a few labeled volumes has exactly three annotated orthogonal slices (sagittal, coronal, axial). The method combines a CLIP-based semantic-guided auxiliary learning module, which injects text-embedding-derived cross-modal parameters into three V-Net branches, with a triple-view disparity training strategy that supervises each branch with two of the three slice labels and cross-supervises unlabeled volumes with uncertainty-filtered pseudo-labels from the other two branches. Experiments on LA2018, KiTS19, and LiTS report state-of-the-art Dice and boundary metrics against seven semi-supervised baselines, together with ablations of the components, prompt templates, annotation geometry, and the loss weight alpha. Code is released.","tokens_in":13433,"tokens_out":7013,"duration_ms":52427,"significance":"If the reported results were robust, the paper would make a useful contribution to reducing annotation cost: the clinical protocol of annotating only three slices is plausible, the CLIP-adapter mechanism is simple to reproduce, and the triple-view co-training idea is a reasonable extension of orthogonal-slice methods such as Desco. The release of code is valuable. However, the headline comparison does not currently isolate the algorithmic contribution: Tables 1-3 give SGTC an extra annotated slice relative to every baseline, and the matched-budget comparison in Table 5 shows only a 0.2 Dice gain over BCP without error bars or significance testing. The paper's own limitation statement admits sensitivity to the choice of annotated slices, which is central to the proposed clinical protocol. I therefore see the contribution as promising but not yet established.","major_comments":[{"comment":"The main SOTA comparison is confounded by annotation budget. In Tables 1-3 all baselines and the SGTC(Dual) ablation use CA (two orthogonal slices), while SGTC(Ours) uses SCA (three orthogonal slices). Since the text reports improvements such as +6.3% Dice over the second-best method on LA2018, the gain may simply reflect the extra slice of supervision. Internal evidence supports this concern: on LA2018, SGTC(Dual) at the same two-slice budget scores 0.739 Dice versus 0.784 for BCP, so SGTC is worse than BCP until it receives the third slice. The central claim should be re-tested with all baselines run at the SCA budget, or re-framed as a comparison of annotation protocols rather than of algorithms.","section":"Comparison Experiments (Tables 1-3)"},{"comment":"The controlled comparison is underpowered. Table 5, the only place where all methods use the same CAC or SCA strategy on KiTS19, reports SGTC at 0.911/0.915 Dice versus BCP at 0.909/0.913, a 0.2-point gap, with no error bars or pairwise significance test. The accompanying sentence states that the performance gains are due to the more effective triple-view disparity training strategy, which is stronger than the evidence in the table. Please provide repeated-run statistics (e.g., mean plus/minus standard deviation over multiple seeds and a paired test) or soften the claim to competitive.","section":"Ablation Study and Analysis (Table 5)"},{"comment":"In the LiTS comparison, the SGTC (Dual) row is reported with 80 unlabeled scans while every other row uses 90 unlabeled scans. This unequal unlabeled pool makes the row incomparable and undermines the cross-dataset consistency of the comparison; it also matters for semi-supervised methods, where the size of the unlabeled pool influences performance. Please correct the count or explain the discrepancy, and rerun the comparison with matched unlabeled data.","section":"Comparison Results on LiTS (Table 3)"},{"comment":"The slice-selection protocol assumes access to foreground location information. The Implementation Details state that the three selected slices should contain the foreground area of the segmentation target and should be chosen as close to the center position as possible in all three planes, and the Conclusion admits performance degradation when the selected slices contain limited foreground information. Because slice selection is informed by target location, the claimed clinical scenario, in which radiologists just need to annotate three orthogonal slices, is not fully demonstrated. I ask for an experiment that varies slice positions (e.g., central versus off-center) and reports Dice, so readers can assess sensitivity to this assumption.","section":"Implementation Details / Conclusion"}],"minor_comments":[{"comment":"Equation (6) contains a malformed normalization expression; please rewrite it cleanly and define all variables (H, W, D, k) at first use.","section":"Methodology, Eq. (6)"},{"comment":"The dynamic alpha schedule is under-specified: increased every 150 iterations does not state the increment or update rule. Since Table 7's main result depends on this schedule, please give the exact schedule (e.g., linear or step, from alpha = 0.1 to which maximum).","section":"Implementation Details, Eq. (10)"},{"comment":"The t-SNE description is confusing: it says we trained five networks with just one single annotated slice per volume, then lists three orthogonal slices for training s, c, and a1, and three parallel slices for training a1, a2, and a3; please clarify the setup and the labels in the figure.","section":"Ablation Study and Analysis, Figure 7"},{"comment":"The LiTS dataset is cited to Heller et al. 2019 in this paragraph, whereas the Datasets section cites Bilic et al. 2023; please correct the citation.","section":"Comparison Experiments, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is reasonable and the paper is clearly written, but the experimental framing overstates the result. The revision should be judged on whether the authors can provide matched-budget comparisons with significance statistics across all datasets, or alternatively reframe the contribution as a comparison of annotation protocols plus a competitive method. I would not recommend rejection on the current evidence, but the present major-superiority claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: this paper is a reasonable engineering combination of known pieces—three orthogonal annotated slices, CLIP text embeddings through an adapter, and triplet co-training with uncertainty-weighted pseudo-labels. That combination is new in the sparse-annotation literature, and the code is out. The ablation study is actually decent: each component helps, the prompt template matters modestly, and three orthogonal slices beat two orthogonal or three parallel ones. The authors also acknowledge the obvious limitation that slice selection has to be informed by foreground location.\n\nThe soft spot is the one the stress-test note flags, and it's real. Tables 1–3 give SGTC three annotated slices per labeled volume while all baselines, including Desco, get only two. So the 2–6 point Dice gaps mostly measure the extra annotation budget, not the algorithm. The only matched-budget comparison, Table 5 on KiTS19, shows SGTC at 0.915 Dice vs BCP's 0.913 with no error bars or significance test. That's a rounding-level edge, and it's the load-bearing evidence for 'triple-view disparity training' being better than a strong two-view baseline. There's also the unexplained LiTS inconsistency where SGTC(Dual) is listed with 80 unlabeled volumes while every other row uses 90. That should have been caught.\n\nNone of this is fatal—the method is plausible and the paper is honest about the slice-selection dependence—but the headline claim is overreach. The experiment section needs a matched-budget comparison for all datasets, standard deviations, and ideally a significance test or at least a repeated-run table. The annotation protocol also needs a clearer statement about who picks the slices and what happens when the organ doesn't intersect all three planes cleanly.\n\nWho should read it: anyone working on barely-supervised 3D segmentation; the three-orthogonal-slice setup is a practical protocol worth citing. I'd send it to review, but with a request for the controlled experiments and a fix to the LiTS table. A serious editor shouldn't desk-reject this; it just needs the evidence to match the claim.","headline":"A sensible three-slice annotation protocol with CLIP guidance, but the claimed SOTA edge over two-slice methods is not yet proven because the main comparison is budget-unfair.","tokens_in":13952,"tokens_out":2233,"would_cite":true,"duration_ms":20377,"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":"SGTC claims that three orthogonally chosen annotated slices per volume, plus CLIP-guided semantic conditioning and triplet co-training, are enough to outperform existing semi-supervised medical image segmentation methods on three public…","keywords":["semi-supervised learning","medical image segmentation","sparse annotation","orthogonal slices","co-training","CLIP","pseudo-labels","3D segmentation"],"falsifier":"Take the LA2018 or KiTS19 training set and replace the center-based slice selection with three randomly chosen orthogonal slices per volume, keeping all other training settings identical. If the Dice and surface-distance margins over BCP and Desco shrink to near zero or reverse, the advantage depends on informed slice selection rather than on the triplet co-training mechanism itself.","tokens_in":1778,"feed_emoji":"🩻","tokens_out":1997,"duration_ms":44957,"temperature":0.7,"pith_summary":"The paper proposes a framework, SGTC, that aims to make semi-supervised medical image segmentation practical with extremely sparse annotation: only three slices per labeled volume, one from each orthogonal plane (sagittal, coronal, axial). The authors report that on LA2018, KiTS19, and LiTS, using 10% labeled cases, SGTC beats recent sparse- and dense-annotation methods on Dice, Jaccard, Hausdorff distance, and average surface distance. If the claim holds, radiologists would annotate a handful of slices instead of entire volumes while still getting state-of-the-art segmentations, especially at weak boundaries. The contribution is presented as two coupled mechanisms: semantic guidance from a frozen CLIP text encoder, and a triple-view co-training scheme that keeps three sub-networks distinct while exchanging pseudo-labels.","feed_headline":"Three orthogonal slices per volume beat full-volume labels","feed_subtitle":"SGTC co-trains three views with CLIP semantics, topping Dice on LA2018, KiTS19, and LiTS at 10% labeled data","key_machinery":"The mechanism is a two-part architecture. First, triple-view disparity training: three V-Net sub-networks, $F_s$, $F_c$, $F_a$, are trained so that each sees two of the three orthogonal annotated slices as its own supervision, and on unlabeled volumes each receives pseudo-labels from the other two, filtered by Monte Carlo dropout uncertainty. Second, semantic-guided auxiliary learning: a frozen CLIP text encoder plus an adapter produces a text embedding $w$, which is concatenated with the global image feature and passed through an MLP to yield cross-modal parameters $\\gamma$ that are element-wise added to the features before the classification layer. This mechanism carries the argument by injecting semantic context to sharpen boundaries and by preserving the spatial disparity needed for meaningful co-training.","core_discovery":"The paper's central claim is that a triplet of 3D segmentation networks, each supervised by a different pair of the three orthogonal slice labels and cross-supervised on unlabeled volumes by the other two networks' uncertainty-filtered predictions, can produce segmentations that surpass previous semi-supervised methods under sparse annotation. The semantic branch takes a medical text prompt through a frozen CLIP encoder and a small adapter, concatenates the resulting text embedding with the image feature, and uses an MLP to produce per-network conditioning parameters that are added into the features before classification. The authors assert this combination yields semantic-aware, fine-granular segmentation, with the largest reported gains over the second-best method on LA2018 (6.3% Dice) and consistent improvements on the CT datasets.","pith_inferences":["The paper's comparison with BCP at matched annotation strategy in Table 5 shows a small gap (0.915 vs 0.913 on KiTS19), which suggests the triple-view scheme may matter less than the extra third slice; a fair test would be to give BCP the same three orthogonal slices and compare.","Because CLIP was trained on natural images rather than medical data, the text branch may act more as a generic feature-modulation prior than as true medical semantic knowledge; replacing the CLIP embedding with a fixed random projection would test this.","The method assumes central slices that contain the target organ; for lesions near the periphery or for organs with high anatomical variability, the informed-slice requirement could be a practical bottleneck, and an extension could select slices automatically based on foreground likelihood.","The semantic conditioning is reported to improve boundary metrics, so a natural next check is to see whether the gains concentrate in Hausdorff distance and average surface distance rather than Dice."],"forward_implications":["Radiologists could label three slices per volume instead of slice-by-slice full volumes, cutting annotation cost while reportedly improving segmentation quality on the tested organs.","The triplet co-training design generalizes across MR and CT modalities because it relies only on the standard three anatomical planes and a 3D backbone.","Pseudo-label quality is claimed to improve enough that the framework can learn effectively from about 90% unlabeled volumes per batch setting.","The semantic branch works with a frozen natural-image CLIP encoder, suggesting text-guided conditioning can be added to existing segmentation networks without retraining the text model.","The reported ablation shows both components contribute, and the dynamic loss weight $\\alpha$ stabilizes training better than fixed values."],"supporting_citations":[{"why":"Desco is the primary orthogonal-annotation baseline that SGTC extends and compares against, supplying the two-slice sparse annotation setup and dataset splits.","marker":"Cai et al. 2023"},{"why":"BCP is the strongest pseudo-label baseline in the comparisons and the key method SGTC must beat on all three datasets.","marker":"Bai et al. 2023"},{"why":"CLIP provides the frozen text encoder whose embeddings are the source of semantic guidance in the SGAL branch.","marker":"Radford et al. 2021"},{"why":"UA-MT supplies the Monte Carlo dropout uncertainty estimation used to filter pseudo-labels and the LA2018 training/test split protocol.","marker":"Yu et al. 2019"},{"why":"V-Net is the 3D backbone architecture for all three sub-networks and is where the semantic conditioning parameters are injected.","marker":"Milletari, Navab, and Ahmadi 2016"},{"why":"LA2018 is the left atrium MRI benchmark used in the first main comparison and ablation experiments.","marker":"Xiong et al. 2021"},{"why":"KiTS19 is the kidney tumor CT dataset used in the second main comparison and the component ablations.","marker":"Heller et al. 2019"},{"why":"LiTS is the liver and liver tumor CT benchmark, and its split follows the protocol used in Desco.","marker":"Bilic et al. 2023"}],"fun_headline_variants":["CLIP-guided triplet co-training needs just 3 slices per volume","3 slices per volume: CLIP semantics make co-training win","Semantic-guided triplet co-training: sparse labels, strong segmentation","Three orthogonal slices plus CLIP edge out full-volume labels","SGTC: sparse triple-view co-training ups Dice with CLIP"],"cache_read_input_tokens":16128,"weakest_assumption_plain":"The reported gains assume a radiologist can select, for each labeled volume, three orthogonal slices that lie near the target organ's center and contain substantial foreground; the paper itself notes in the Conclusion that performance degrades when the selected slices contain limited foreground information.","fun_headline_variants_meta":{"raw":{"variants":["CLIP-guided triplet co-training needs just 3 slices per volume","3 slices per volume: CLIP semantics make co-training win","Semantic-guided triplet co-training: sparse labels, strong segmentation","Three orthogonal slices plus CLIP edge out full-volume labels","SGTC: sparse triple-view co-training ups Dice with CLIP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001164,"raw_usage":{"total_tokens":4818,"prompt_tokens":946,"completion_tokens":3872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":3785}},"tokens_in":562,"tokens_out":3872,"duration_ms":25448,"temperature":1.0,"reasoning_tokens":3785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:20:29.097344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the LA2018 or KiTS19 training set and replace the center-based slice selection with three randomly chosen orthogonal slices per volume, keeping all other training settings identical. If the Dice and surface-distance margins over BCP and Desco shrink to near zero or reverse, the advantage depends on informed slice selection rather than on the triplet co-training mechanism itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Desco is the primary orthogonal-annotation baseline that SGTC extends and compares against, supplying the two-slice sparse annotation setup and dataset splits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BCP is the strongest pseudo-label baseline in the comparisons and the key method SGTC must beat on all three datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"V-Net is the 3D backbone architecture for all three sub-networks and is where the semantic conditioning parameters are injected."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LA2018 is the left atrium MRI benchmark used in the first main comparison and ablation experiments."},{"cited_title":"B.; Vorontsov, E.; Ben-Cohen, A.; Kaissis, G.; Szeskin, A.; Jacobs, C.; Mamani, G","cited_arxiv_id":null,"evidence_quote":"LiTS is the liver and liver tumor CT benchmark, and its split follows the protocol used in Desco."}],"review_version":1}