{"id":"c006d4f8-bec8-464e-83a9-04f580aab0f8","arxiv_id":"2507.00519","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TopoNet combines snake convolutions, RGB-D fusion, and topology-preserving losses to improve laparoscopic liver landmark detection accuracy and speed.","lead":"This paper introduces TopoNet, a deep learning model that detects liver landmarks in laparoscopic video frames using both RGB images and estimated depth maps, with losses that enforce topological consistency. The authors report higher accuracy and up to four times faster inference than twelve competing methods on two surgical datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central SOTA claim is undercut by an unexplained 5.40-point DSC discrepancy between Table 1 and Table 2 for the same L3D evaluation set.","rationale":"The paper's central claim is an empirical state-of-the-art result. For that claim to hold, the headline comparison in Table 1 must be internally consistent with the ablation in Table 2. It is not: the full model is 5.40 DSC points lower in Table 2 on what is described as the same L3D evaluation set, and this puts TopoNet below the second-best method D2GPLand. The reader's identified weak point concerns the reliability of frozen AdelaiDepth depth maps; while a plausible risk, it is secondary because even if depth maps are imperfect, the empirical comparison could still be valid. The table conflict, by contrast, directly invalidates the comparison unless resolved. It could be a typo or different split, so rejection is not mandatory, but the current manuscript does not support the SOTA claim as written. The conditional verdict remains appropriate, with the added explicit condition that the authors reconcile Table 1 and Table 2 and release code or exact split definitions.","tokens_in":7934,"tokens_out":4253,"duration_ms":51393,"concrete_test":"Obtain the released code and exact L3D split, then train the full TopoNet under both the Table 1 and Table 2 protocols and compare on the identical test set. Ask the authors to state explicitly which subset Table 2 uses and to report scores from the best checkpoint of Table 1 plus three seeds. If the full model cannot reproduce 65.19 DSC on the L3D test set, the central claim is unsupported; if it can, the Table 2 numbers must be relabeled as a different evaluation subset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3/Table 1 reports TopoNet at 65.19 DSC / 50.56 IoU / 28.07 Assd on L3D, beating D2GPLand (63.52/48.68/59.38) and eleven other methods. Section 3.4/Table 2, described as using the evaluation set of L3D, lists the full TopoNet at 59.79 DSC / 47.38 IoU / 29.27 Assd. DSC differs by 5.40 points and IoU by 3.18 points. Taken literally, the full model in the ablation is below D2GPLand's 63.52 DSC, so the SOTA claim would fail. The paper offers no split definition or reconciliation; if Table 2 uses a validation subset rather than the Table 1 test set, that is not stated and the comparison protocol is not held fixed across tables. Because every downstream claim (depth fusion, topological losses, efficiency) is evidence for the headline, this unexplained discrepancy is the most load-bearing internal inconsistency.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TopoNet, a framework for laparoscopic liver landmark detection that combines a snake-convolution and ResNet dual-path encoder to extract depth-informed topological features and RGB texture features, a boundary-aware topological fusion (BTF) module that merges the two modalities while emphasizing edge regions, and a topological constraint loss comprising a multi-class center-line (clDice-style) loss and a persistent-homology-based matching loss (Eqs. 4-9). On the L3D and P2ILF datasets, the authors report state-of-the-art DSC/IoU/Assd across 12 baselines (Table 1), with the headline result of 65.19 DSC on L3D versus 63.52 for the prior best method D2GPLand, a large reduction of average symmetric surface distance (28.07 vs 59.38), roughly 4x faster inference (86.43 ms vs 297.93 ms), and about half the FLOPs. Ablation studies (Table 2, Fig. 5) attribute gains to the BTF module and to each topological loss component. The central problem identified in this review is that the full model's L3D scores in the ablation table (59.79 DSC, 47.38 IoU, 29.27 Assd) differ from the headline test-set scores (65.19 DSC, 50.56 IoU, 28.07 Assd) without any stated explanation of the evaluation protocol.","tokens_in":8102,"tokens_out":11306,"duration_ms":121113,"significance":"If the reported numbers are reproducible, the paper would make a solid contribution to intraoperative landmark detection: the loss formulations in Eqs. (4)-(9) are concretely specified, the design motivations (snake convolutions for tubular anatomy, depth-informed topology, explicit RGB-D fusion) are sensible, and the efficiency gains over D2GPLand are substantial and clinically relevant. The paper also reports a more complete efficiency profile (inference speed and GFLOPs) than is typical for this literature, and the component ablations are informative as far as they go. However, the unresolved discrepancy between Table 1 and Table 2, the absence of error bars or significance tests on margins of about 1-2 DSC points, and the self-defined random P2ILF split mean that the state-of-the-art claim is not verifiable from the manuscript as written. These issues appear fixable, and the underlying approach remains promising.","major_comments":[{"comment":"The full TopoNet model is reported with L3D DSC/IoU/Assd of 65.19/50.56/28.07 in Table 1 but 59.79/47.38/29.27 in Table 2, although Section 3.4 states that the ablations use \"the evaluation set of L3D.\" If \"evaluation set\" is the same test set used for Table 1, the two tables are mutually contradictory, and the state-of-the-art claim in Section 3.3 fails, since the Table 2 full model would rank below D2GPLand (63.52 DSC). If the ablation instead uses a validation subset or another split, that must be stated explicitly, baseline methods must be re-run on the same split, and the protocol difference must be reconciled; as written, no downstream claim (depth fusion, topological losses, or efficiency) can be evaluated against a fixed comparison protocol.","section":"3.3/Table 1 vs 3.4/Table 2"},{"comment":"Every metric in Tables 1 and 2 is a single-run point estimate with no error bars, no number of repeated runs, and no statistical significance test. The headline margins over D2GPLand (1.67 DSC on L3D, 0.81 DSC on P2ILF) are smaller than or comparable to the unexplained 5.40-point DSC gap between the two reported evaluations of the full TopoNet model, so run-to-run variability alone could account for the reported gains. Please report mean +/- standard deviation over at least three runs with different seeds and a paired significance test for the TopoNet-versus-D2GPLand comparison on both datasets.","section":"3.3 and 3.4"},{"comment":"The P2ILF comparison is conducted on a self-defined random split (124 training images and 43 test images) with no seed stated and no repeated splits, and the paper does not state whether the 12 baseline numbers in the P2ILF column were obtained by re-running those methods under this identical split or quoted from prior publications that used different splits. Unless every baseline was re-trained and evaluated under exactly this split with the same protocol, the P2ILF column of Table 1 is not a controlled comparison, and the reported 0.81 DSC improvement cannot be interpreted; if the baselines were re-run, the training details for each one must be reported.","section":"3.1 and Table 1 (P2ILF)"},{"comment":"The entire depth pathway depends on depth maps produced by the frozen monocular estimator AdelaiDepth, which is not validated on laparoscopic imagery, yet the paper provides no analysis of depth quality or of RGB-depth alignment and no ablation that removes the depth pathway (all Table 2 variants still use depth, with or without BTF). As a concrete test, please report an RGB-only variant (STA blocks removed and the BTF module fed with R_i only) together with a brief depth-quality evaluation, so that the gains attributed to \"depth-informed topological structures\" can be separated from the mere addition of a second modality and the reader can assess the risk of failure on glossy, deformable liver surfaces.","section":"2.1, 3.2, and Table 2"}],"minor_comments":[{"comment":"The sentence \"the other ones utilize only the former fused feature F_{i-1} for residual learning\" is ambiguous: it is unclear whether BTF modules 2-5 still receive R_i and D_i as inputs or only F_{i-1}; the figure suggests both modalities remain inputs, so the text should describe the data flow unambiguously.","section":"2.2 and Eq. (3)"},{"comment":"The sentence describing the P2ILF split (\"Since only the training set can be available, we randomly select 124 images for training and the remaining 43 images for testing\") is confusing; please state clearly that 16 of the 183 images form a held-out set with unavailable labels and that the reported experiments re-split the 167-image training set into 124/43, and report the random seed.","section":"3.1"},{"comment":"The caption says that a broken landmark is highlighted by a yellow circle and two topological false positives by red circles, but it does not state in which panels these circles appear; readers cannot tell whether the broken landmark is in the D2GPLand prediction (c) or elsewhere.","section":"Fig. 1 caption"},{"comment":"The claim that the loss ensures \"homotopy equivalence between predictions and labels\" is stronger than what Eqs. (6)-(8) implement, which is a 0-dimensional persistent-homology (connected-component) matching loss plus center-line connectivity; please rephrase to describe the actual topological consistency enforced.","section":"2.3 and Abstract"},{"comment":"The backbone ablation is presented only as a figure with no numerical values; please provide a table (or add values to the caption) so that the ResNet-34 choice is quantitatively justified, and state which backbone variants were compared.","section":"3.4 and Fig. 5"},{"comment":"Minor wording and reporting issues: \"the inference speed of TopoNet outperform all compared models\" should read \"outperforms,\" \"GLOPs\" should be \"GFLOPs,\" and the paper should state the input resolution used for the speed and GFLOPs measurements.","section":"3.3"},{"comment":"The paper has no related-work section and does not compare against existing topology-preserving losses for segmentation (e.g., the TopologyLoss of Hu et al., NeurIPS 2019, or other Betti-matching losses); adding such a comparison would clarify the novelty of L_per and help position the contribution.","section":"1 and Related Work"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the experimental protocol is largely inherited from the authors' own prior MICCAI 2024 paper (D2GPLand), and the baseline numbers appear to come from that internal protocol; combined with the Table 1/Table 2 split ambiguity and the unreported P2ILF random seed, the current evidence is reproducible only by the authors themselves. I suggest making the release of exact split file lists, code or pretrained checkpoints, and per-run results a condition of acceptance, and asking that the discrepancy between Tables 1 and 2 be resolved in the next revision. The fit with the journal's scope is appropriate, and I see no novelty-disclosure concern, but the self-comparison pattern (the strongest baseline is the authors' own prior work, evaluated under their own protocol) warrants a careful check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TopoNet is a sensible integration of known components—snake convolutions, multi-class clDice, persistence-based loss—for a real clinical problem, and the efficiency gains are credible. The central claim is that it beats 12 methods on L3D and P2ILF, but the paper currently has an unexplained 5.4-point DSC gap between its main table and its ablation table for the same model on L3D. That needs to be resolved before the SOTA claim can be taken seriously.\n\nTo the paper's credit, the architecture is well-motivated. The snake-CNN dual-path encoder is a reasonable way to use depth for topology while keeping RGB texture, and the boundary-aware fusion module is a plausible mechanism. The center-line loss extends clDice to multiple classes, and the persistence loss directly targets the topological false positives that inflate Assd. The 31-pixel Assd improvement over D2GPLand is large, but the authors give a mechanism: outliers that would otherwise create disconnected false positives are penalized. The efficiency numbers, roughly 3.4x faster and half the GFLOPs of D2GPLand, are also believable.\n\nThe soft spots are proportional. The Table 1/Table 2 discrepancy is the biggest. If Table 2 uses a validation subset, the text must say so; if it uses the test set, the model is below the D2GPLand result it claims to beat. Either way, the protocol is not held fixed across tables. Also, there are no error bars or multi-seed statistics, so all point estimates are fragile. The code is promised but absent, and the behavior of the frozen AdelaiDepth depth estimates on glossy, deformable liver is never analyzed. Finally, 'homotopy equivalence' overstates what a persistence loss plus clDice enforces; these losses encourage topological consistency but don't guarantee it.\n\nBottom line: the task is practical, the contribution is real but incremental, and the paper deserves a serious referee. I would recommend conditional acceptance only if the authors reconcile the two tables, release the code, and run the ablations with at least three seeds. Until then, I would not cite the SOTA claim in my own work.","headline":"A plausible integration of known topology-preserving losses, but the unexplained 5.4-point DSC gap between the main table and the ablation table currently undercuts the SOTA claim.","tokens_in":8669,"tokens_out":3270,"would_cite":false,"duration_ms":41165,"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":"Explicit topological supervision lifts laparoscopic liver landmark detection to state-of-the-art accuracy while cutting inference time to about a third of the closest rival's.","keywords":["topology-constrained learning","laparoscopic liver landmark detection","persistent homology loss","center-line constraint loss","RGB-D fusion","dynamic snake convolution","depth estimation","surgical augmented reality"],"falsifier":"Replace the estimated depth maps in the frozen AdelaiDepth stream with random noise while keeping the same training pipeline; if DSC, IoU, and Assd stay near the reported values, the depth-informed topology claim is not load-bearing, and if the gains disappear the method depends on depth quality. A second check is to measure AdelaiDepth's depth error against the L3D keyframes and see whether frames with poor depth are exactly the frames where TopoNet's Assd advantage shrinks.","tokens_in":7708,"feed_emoji":"🩺","tokens_out":5059,"duration_ms":50656,"temperature":0.7,"pith_summary":"Liver landmarks are thin, winding anatomical structures inside the abdomen, and this paper's claim is that a detector trained to respect their topology—their connectedness and branching—rather than only their pixel overlap will find them reliably even when instruments, blood, or similar tissue are in the way. To test that, the authors build TopoNet, a two-path network: a ResNet reads the RGB image for texture, while stacking snake-convolution blocks read estimated depth maps for the landmarks' tubular shape. A boundary-aware fusion module merges the two streams, and the training objective adds a center-line constraint and a persistent-homology loss that enforce homotopy equivalence between prediction and ground truth. On the L3D and P2ILF benchmarks the paper reports the best numbers among twelve compared methods, with the largest gain in average symmetric surface distance, and about a 3.4 times faster inference than its closest competitor.","feed_headline":"Topology supervision halves liver landmark detection error distance","feed_subtitle":"Adding depth-guided topological shape constraints to a dual-path encoder beats 12 prior methods and runs ~3.4x faster than the runner-up.","key_machinery":"The load-bearing mechanism is the topological constraint loss, which pairs a multi-class center-line constraint (an extension of clDice) with a persistent-homology loss. The persistence loss computes matched and unmatched connected components between prediction and ground-truth barcodes; matched components are penalized for birth–death interval disagreement, unmatched predicted components are penalized for their persistence length, so predictions are pulled toward homotopy equivalence with labels. The architecture that feeds this loss is the snake-CNN dual-path encoder: STA blocks apply dynamic snake convolution to depth maps to capture tortuous tubular geometry, a ResNet-34 extracts RGB texture, and boundary-aware topological fusion modules combine the streams with attention and a boundary-enhancement residual.","core_discovery":"The paper's central result is that explicit topological supervision changes which errors a landmark detector makes. On L3D, TopoNet improves Dice from 63.52 to 65.19 and IoU from 48.68 to 50.56 over D2GPLand, while average symmetric surface distance drops from 59.38 to 28.07 pixels; on P2ILF the gains are 0.81 in Dice, 1.01 in IoU, and 8.57 pixels in Assd. Inference time falls from 297.93 ms to 86.43 ms and computation from 572.85 to 276.99 GFLOPs. The authors attribute the large Assd improvement to the persistence loss suppressing topological false positives caused by surgical instruments and blood, which sit far from any true landmark pixel.","pith_inferences":["In my reading, the depth maps are the fragile link: the gains are only as reliable as the frozen AdelaiDepth estimates, and the paper reports no validation of depth quality on the glossy, deforming liver surface; I would expect accuracy to degrade if the depth stream is perturbed.","The persistence loss is stated generically, so it could be lifted to vessel, nerve, or duct segmentation tasks where the target topology is known, with the same matched/unmatched barcode penalty replacing pixelwise losses.","A direct comparison against an RGB-only TopoNet variant (no STA depth path) would isolate how much of the 1.67-point DSC gain comes from depth geometry versus from the topological losses themselves; the ablations currently remove losses and fusion but not the depth pathway."],"forward_implications":["TopoNet reports the best DSC, IoU, and Assd on both L3D and P2ILF compared with twelve prior methods, including transformer- and SAM-based segmenters.","The Assd drop from 59.38 to 28.07 pixels on L3D suggests that far-off false positives from occluding instruments and blood are largely removed rather than merely shrunk.","At 86.43 ms per frame and 276.99 GFLOPs, the model is light enough that intraoperative landmark localization for AR-guided liver surgery is feasible on current GPU hardware.","The center-line and persistence losses are direct extensions of clDice and persistent homology, so the same training recipe can be applied to other tubular structures without architectural changes."],"supporting_citations":[{"why":"Provides the L3D dataset, the D2GPLand depth-geometric-prior baseline that TopoNet builds on and beats, and the evaluation metrics and settings.","marker":"[16]"},{"why":"Supplies the frozen monocular depth estimator whose depth maps feed the snake-CNN dual-path encoder.","marker":"[25]"},{"why":"Contributes dynamic snake convolution, which the STA blocks use to capture tortuous tubular landmark geometry.","marker":"[17]"},{"why":"Defines clDice, the topology-preserving tubular loss that the paper extends to multi-class center-line supervision.","marker":"[20]"},{"why":"Provides the efficient barcode computation used to match connected components between prediction and ground truth for the persistence loss.","marker":"[22]"},{"why":"Supplies the persistent homology theory that the topological persistence loss is built on.","marker":"[7]"},{"why":"Provides the P2ILF dataset and the augmented-reality challenge context; the paper evaluates on it as a second benchmark.","marker":"[1]"},{"why":"Provides the ResNet-34 backbone that extracts the RGB texture features in the dual-path encoder.","marker":"[10]"}],"fun_headline_variants":["TopoNet cuts liver landmark error distance by half","Topology-aware learning halves liver landmark error","Faster and sharper: TopoNet for liver landmarks","Depth-aware topology boosts liver landmark detection","TopoNet: topology constraints improve liver surgery guidance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth maps fed to the snake-CNN path come from a frozen monocular depth estimator, and the method's gains assume those depth estimates are accurate and aligned with the RGB frames on the glossy, deformable liver surface.","fun_headline_variants_meta":{"raw":{"variants":["TopoNet cuts liver landmark error distance by half","Topology-aware learning halves liver landmark error","Faster and sharper: TopoNet for liver landmarks","Depth-aware topology boosts liver landmark detection","TopoNet: topology constraints improve liver surgery guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1449,"prompt_tokens":894,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":483}},"tokens_in":510,"tokens_out":555,"duration_ms":6335,"temperature":1.0,"reasoning_tokens":483,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:13:01.830944+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the estimated depth maps in the frozen AdelaiDepth stream with random noise while keeping the same training pipeline; if DSC, IoU, and Assd stay near the reported values, the depth-informed topology claim is not load-bearing, and if the gains disappear the method depends on depth quality. A second check is to measure AdelaiDepth's depth error against the L3D keyframes and see whether frames with poor depth are exactly the frames where TopoNet's Assd advantage shrinks.","supporting_citations":[{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"Provides the L3D dataset, the D2GPLand depth-geometric-prior baseline that TopoNet builds on and beats, and the evaluation metrics and settings."},{"cited_title":"IEEE TPAMI (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen monocular depth estimator whose depth maps feed the snake-CNN dual-path encoder."},{"cited_title":"In: IEEE ICCV","cited_arxiv_id":null,"evidence_quote":"Contributes dynamic snake convolution, which the STA blocks use to capture tortuous tubular landmark geometry."},{"cited_title":"In: IEEE CVPR","cited_arxiv_id":null,"evidence_quote":"Defines clDice, the topology-preserving tubular loss that the paper extends to multi-class center-line supervision."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the efficient barcode computation used to match connected components between prediction and ground truth for the persistence loss."},{"cited_title":"MedIA99, 103371 (2025)","cited_arxiv_id":null,"evidence_quote":"Provides the P2ILF dataset and the augmented-reality challenge context; the paper evaluates on it as a second benchmark."},{"cited_title":"In: IEEE CVPR","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet-34 backbone that extracts the RGB texture features in the dual-path encoder."}],"review_version":1}