{"id":"50e48f5f-eb3e-463d-b4ca-38f5b8f22b76","arxiv_id":"2412.20608","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new convolution layer uses persistent homology on feature maps to guide adaptive kernel offsets, improving topological consistency in medical image segmentation.","lead":"This paper introduces a convolutional layer whose sampling positions are steered by topological priors computed with persistent homology, aiming to preserve connectivity in medical image segmentation. It reports improved topology metrics across retinal vessel, colon cancer cell, and neuron electron microscopy datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal role of the PH-derived topological prior is untested: no ablation compares the TPG against a non-topological prior, and the fixed tau_0/psi choices are unspecified, so the claimed mechanism remains unsupported.","rationale":"The reader's weakest assumption identified the unspecified tau_0 and pooling operator as key implementation gaps. I agree those are real, but the more decisive issue is that the paper never demonstrates that the PH-derived prior is the cause of the topological improvements. The ablations in Tab. 4 only remove post-processing steps of PH outputs; they do not compare against a prior that is not based on persistent homology at all. If a random or edge-based prior confers the same topological benefit, then the observed gains would arise from the conformable architecture rather than from topological reasoning, undermining the central claim in the abstract and Section 4. The text in Sec. 4.1 does not specify psi or tau_0, and the experiments do not include a sensitivity analysis, so a fixed preset threshold may be a hidden critical choice that does not transfer across datasets or architectures. These issues do not disprove the method, but they leave the claim that topological awareness is the active mechanism conditional on additional evidence. Therefore the existing CONDITIONAL verdict remains appropriate without modification.","tokens_in":14351,"tokens_out":5290,"duration_ms":56550,"concrete_test":"On the CHASE UNet+Conform setup, replace the PH prior phi_pr with three fixed non-topological priors, keeping all downstream components (Gaussian dilation, aggregation, offset generator) unchanged: (a) a random binary map with the same density as the filtered generators, (b) a Sobel edge-magnitude map, (c) the channel-pooled feature map phi_pooled directly, without PH or filtering. Compare error_beta0, error_beta1, and clDice. Additionally, sweep tau_0 over at least five values spanning the persistence distribution and report the same topological metrics. If any non-PH prior performs within one standard deviation of the full TPG, or if the topology metrics swing by more than one standard deviation across the tau_0 sweep, then the PH-specific prior and the fixed threshold are not load-bearing for the reported topology preservation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that persistent-homology-derived topological priors (TPG) guide the offset generator to focus on topologically significant regions, thereby preserving connectivity. The experimental support relies on (i) comparisons against deformable convolution and DSC, whose offset generators consume the raw feature map rather than phi_post, and (ii) ablations in Tab. 4 that remove filtration, Gaussian dilation, or aggregation while always retaining the PH-generated generator coordinates as the prior. Consequently, the active ingredient is never isolated: the conformable architecture itself (extra conv branch, aggregation, dilation) could be driving the topological gains, with the PH map merely acting as one of many possible inputs. This gap is compounded by the unspecified pooling operator psi and threshold tau_0 in Sec. 4.1; since tau_0 is set without stated rationale and never swept, robustness to this parameter is unknown. The decision to use only 0-dimensional persistence also sits uneasily with the reported error_beta1 improvements, suggesting the topological guidance may not directly target holes or loops. Without a no-PH baseline, the causal claim that the PH-derived guidance produces the observed topology preservation is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Conformable Convolution, a convolutional layer whose kernel offsets are guided by a Topological Posterior Generator (TPG) that applies persistent homology to channel-pooled feature maps. The TPG computes birth/death generators in cubical complexes, filters them by a persistence threshold, applies Gaussian dilation, and aggregates the result with the original feature map. The resulting posterior is fed to a small offset-generating network to modulate a second convolution, in the style of deformable convolution. The authors evaluate the layer on three segmentation datasets (CHASE, HT29, ISBI12) using pixel-level metrics and topological/connectivity metrics, and they report ablations of filtration, Gaussian dilation, and feature aggregation, as well as an ablation of the number of Conform blocks.","tokens_in":14594,"tokens_out":4743,"duration_ms":46730,"significance":"If the central claim were established, the paper would contribute a genuinely architecture-agnostic way to inject topological priors into convolutional networks, with potential value for medical segmentation of tubular and networked structures. The idea of using persistent homology to guide kernel offsets is a reasonable departure from both topology-aware loss functions and purely learned deformable convolutions. The paper also makes a positive effort to evaluate with topological metrics beyond Dice, including Betti errors, Euler characteristic, clDice, ARI, and VI, across three datasets. However, the evidence as presented does not isolate the causal role of the PH-derived prior, several key components of the TPG are underspecified, and some reported results are inconsistent with the claim that pixel-level accuracy is not sacrificed. The strengths are the breadth of evaluation and the clarity of the proposed module; the weaknesses are concentrated in missing control experiments and parametric detail.","major_comments":[{"comment":"The causal role of the persistent-homology-derived prior is not isolated. Every ablation in Table 4 retains the PH-generated generator coordinates as the prior and only removes filtration, Gaussian dilation, or aggregation; there is no baseline in which the offset generator consumes a non-topological prior (e.g., the raw feature map, an edge map, or random coordinates) while keeping the rest of the Conformable block fixed. Since the Conformable block adds an extra convolutional branch and a Gaussian-dilation operation beyond the standard deformable convolution, the observed topological gains could plausibly come from those components rather than from the PH guidance. A control experiment with a non-topological prior is needed to support the abstract's claim that PH-guided offsets preferentially focus on topologically significant regions.","section":"Section 4.1, Tables 4 and 5"},{"comment":"The channel pooling operator ψ and the persistence filtering threshold τ0 are never specified, and their sensitivity is not examined. Equation (4) defines φ_pooled = ψ(φ_in) but does not state whether ψ is a mean, max, learned linear combination, or something else. Equation (5) uses an indicator with threshold τ0, but the paper gives no value, no selection criterion, and no sweep over τ0. Because the surviving generator set G_M depends directly on τ0 and on the pooled map, the entire offset-guidance mechanism is not reproducible from the manuscript and its robustness to these free choices is unknown.","section":"Section 4.1, Eqs. (4)-(6)"},{"comment":"There is an unexplained mismatch between the homology dimension used and the metrics reported. The method explicitly tracks 0-dimensional cubes and connected-component generators, yet the paper reports improvements in error β1 and discusses loops and holes (e.g., ISBI12 with 'numerous loops') without explaining how 0-dimensional persistence can directly target 1-dimensional topological features. If the reported β1 improvements are indirect effects of better connected-component preservation, that claim should be stated and supported; otherwise the connection between the TPG mechanism and the β1 results is unsupported.","section":"Section 3, Eqs. (5)-(6), Tables 3-4"},{"comment":"The statement that the method 'does not sacrifice the pixel-wise results' is not supported by the reported statistics. In Table 3, on HT29 the Conform layer has lower Dice (94.6±1.3) than both Deform (95.8±2.1) and DSC (95.8±2.0), and lower AUC (99.1±0.6) than Deform (99.6±0.2); on CHASE, Conform has lower AUC (94.2±0.2) than DSC (95.9±0.2). These are load-bearing for the claim that topological gains come without pixel-level cost. Additionally, Table 5 explicitly states that only the best model is selected and all standard deviations are zero, so the comparison across numbers of Conform layers cannot be assessed statistically. The authors should report all runs and either temper the pixel-level claim or provide a proper model-selection procedure.","section":"Table 3, Table 5, Section 5.2.1"}],"minor_comments":[{"comment":"The word 'Topological' in the abstract is typeset with an erroneous space ('T opological').","section":"Abstract"},{"comment":"The phrase 'the conformal property of our method' appears to be a typo for 'conformable property', since the method is named Conformable Convolution throughout the paper.","section":"Section 5.2.1"},{"comment":"The table numbering and cross-references are inconsistent: Table 1 and Table 3 have nearly identical captions ('Segmentation Performance Compared to SOTA Layers with Adaptive Kernel...'), and the text refers to 'Tab. 3 on ISBI12' for a model comparison that does not appear in Table 3. Please harmonize the table numbers and correct the references.","section":"Tables 1-3 and Section 5.2.1"},{"comment":"Equation (9) defines φ_post = φ_dil ⊙ φ_in + φ_in, while the ablation description in Section 5.2.2 and Eq. (13) use φ_post = φ_dil ⊙ φ_in as the 'blocked aggregation' variant. The text should clarify exactly which operation is being ablated and why removing the residual addition is the intended ablation of 'feature aggregation'.","section":"Eqs. (9) and (13)"},{"comment":"The introduction mentions a new blood-flow-simulation metric in the supplementary materials, and Section 5.1 states that implementation details are reported in the supplement; however, no supplement was provided with the manuscript. Please include the supplement or remove the references to it.","section":"Section 5.1 and Introduction"},{"comment":"The clDice column in Table 2 mixes percentages (e.g., 0.75, 0.73) and apparently unscaled values without a clear legend, making it difficult to compare rows; please use a consistent format.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising, but the current manuscript cannot support the causal claim that the persistent-homology prior produces the observed topology improvements because no baseline without that prior is reported. I would be willing to reconsider after the authors add a no-PH control, specify τ0 and ψ with sensitivity analysis, and fix the statistical reporting in Tables 3 and 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the PH-guided deformable convolution is a legitimate new combination, but the evaluation doesn't prove the causal story. I'd send it to peer review with a request for major revision, not desk reject.\n\nWhat's new: the TPG module computes persistent homology on a pooled feature map and uses the resulting generators to weigh deformable offsets. That specific pipeline I haven't seen before. The layer is architecture-agnostic, and the authors test it on three very different segmentation tasks, which is more than many papers do. The ablations on filtering, Gaussian dilation, and aggregation are the right kind of sanity checks, even if they are incomplete.\n\nWhere it gets soft: the comparisons in Table 3 mostly show that Conform improves Betti/Euler errors on ISBI12 and CHASE, but on HT29 it loses to Deform on Dice (94.6 vs 95.8) and AUC (99.1 vs 99.6), and on CHASE it is below DSC on AUC. Many of the differences are within one standard deviation of three runs. The text says pixel-wise performance is 'not sacrificed,' but the paper's own numbers contradict that on HT29. Table 5 admits best-run selection, and the zero standard deviations there are a real concern.\n\nThe bigger problem is causal. The claim is that the PH prior is what drives the topological gains, but the ablations always keep the TPG coordinates. There is no baseline where the offset generator consumes a random or non-topological prior instead of the PH map. So the extra conv branch, the Gaussian dilation, or the residual aggregation could be doing the work. Also, tau_0 and the pooling operator psi are never specified, and only 0-dimensional persistence is used. The paper does not explain how these choices were made or whether they generalize. The stress-test note on this is spot on.\n\nThere's also no code or supplement details in the version I read, so I can't check the blood-flow simulation metric they mention.\n\nThat said, the core idea is not disproven. The authors know the TDA and segmentation literature, and the paper is coherent on its own terms. With code, a proper sweep of tau_0 and psi, and an ablation that swaps the PH prior for a non-topological one, it could become a solid contribution.\n\nMy recommendation: accept for peer review, but expect major revisions. It deserves referee time, but not a strong accept in this form.","headline":"PH-guided offsets are a genuinely new combination, but the evidence doesn't isolate the mechanism and some claims are contradicted by the paper's own tables.","tokens_in":15080,"tokens_out":2622,"would_cite":false,"duration_ms":27304,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","68U10","92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"Steering convolution offsets with persistent homology preserves connectivity in medical image segmentation without sacrificing pixel accuracy.","keywords":["conformable convolution","topological posterior generator","persistent homology","cubical complexes","deformable convolution","topology-preserving segmentation","medical image segmentation","Betti numbers"],"falsifier":"Retrain the CHASE vessel model with the TPG prior replaced by Gaussian-dilated random generator locations of the same density; if Betti-zero and Euler errors are no worse than with the true topological prior, the specific topological information is not what drives the improvement. Alternatively, sweep $\\tau_0$ from zero to the 90th percentile of persistence values and check whether the reported error reductions depend sharply on the threshold.","tokens_in":14180,"feed_emoji":"🩺","tokens_out":7969,"duration_ms":71995,"temperature":0.7,"pith_summary":"Medical segmentation models can score well on pixel-level accuracy while splitting or merging fine anatomical structures such as retinal vessels and neuron membranes. This paper proposes a convolutional layer, Conformable Convolution, whose kernel sampling positions are steered by a topological prior instead of learned from raw image content alone. The prior is produced by the Topological Posterior Generator, which runs persistent homology on pooled feature maps, keeps only features whose persistence exceeds a threshold, and spreads the surviving generators with Gaussian dilation. On three datasets, retinal vessels, colon cancer cells, and neuron electron microscopy, the layer reduces Betti-number and Euler-characteristic errors compared with deformable and snake-convolution baselines while keeping Dice and AUC broadly comparable. The authors claim the layer is architecture-agnostic and can replace any convolution in a segmentation network.","feed_headline":"Topology-guided kernels cut segmentation connectivity errors","feed_subtitle":"Persistent homology steers kernel offsets to key pixels, improving Betti and Euler errors on three datasets.","key_machinery":"The Topological Posterior Generator is the load-bearing component: it converts a feature map into a spatial prior by channel pooling, cubical persistent homology, persistence-threshold filtering, and Gaussian dilation. It supplies the coordinates, the birth and death locations of 0-dimensional connected components, that the offset-generating convolution should attend to. Conformable Convolution is the second component: a deformable-style convolution in which the offsets are computed as $g(\\text{TPG}(\\phi_{\\text{post}}))$, so the adaptive kernel samples input features at positions $p + p_c + \\Delta\\hat{p}_c$ that are biased toward topologically significant regions.","core_discovery":"The central claim is that persistent homology computed on intermediate feature maps yields a spatial prior that makes adaptive convolution kernels focus on topologically important locations, and that this focus transfers into topologically more faithful segmentation maps. Concretely, the TPG block channel-pools the input feature map, builds cubical complexes, records 0-dimensional persistence pairs, filters generators by a persistence threshold $\\tau_0$, and forms $\\phi_{\\text{post}} = \\phi_{\\text{dil}} \\odot \\phi_{\\text{in}} + \\phi_{\\text{in}}$, where $\\phi_{\\text{dil}}$ is the Gaussian-dilated indicator of surviving generators. Conformable Convolution predicts kernel offsets from $\\phi_{\\text{post}}$ rather than from the raw feature map, so the deformation of the receptive field is biased toward birth and death locations of connected components. The paper reports that this lowers error in $\\beta_0$, $\\beta_1$, and Euler characteristic on CHASE DB1, HT29, and ISBI12 relative to deformable convolution and dynamic snake convolution, while pixel-level metrics stay comparable.","pith_inferences":["A natural extension the authors do not pursue is to feed 1-dimensional persistence generators, loop births and deaths, into the same offset map; ISBI12's holes suggest this could push the reported $\\beta_1$ error below 7.9.","The fixed threshold $\\tau_0$ and the unspecified channel-pooling operator $\\psi$ are hidden design choices; making both adaptive, learned per layer or per dataset, would clarify how much of the effect is generic topological guidance versus dataset-specific tuning.","Because TPG consumes any feature map, the same prior could modulate attention weights in transformers rather than convolution offsets, effectively giving topology-aware attention for segmentation and registration tasks."],"forward_implications":["Inserting one Conformable block at the bottleneck of a U-Net reduces Betti-zero error on CHASE from 26.9 to 21.6 and Euler-characteristic error from 28.5 to 20.6 while keeping Dice near 79 to 80 percent.","Adding the layer to Y-Net raises clDice from 0.76 to 0.79 and cuts Betti-zero error from 27.9 to 21.1 on the same dataset.","Because the module is architecture-agnostic, the same recipe of pooling, persistence, filtering, dilation, and offset prediction should transfer to other encoder-decoder or transformer-based segmentation backbones.","Ablations show that filtration, Gaussian dilation, and feature aggregation each contribute; removing any of them raises topological errors, so all three are needed for the reported gain."],"supporting_citations":[{"why":"Supplies the deformable-convolution mechanism of learnable kernel offsets that Conformable Convolution reuses and redirects.","marker":"[7]"},{"why":"Supplies the persistence-simplification principle used to filter low-persistence generators in the TPG.","marker":"[9]"},{"why":"Dynamic Snake Convolution is the main adaptive-kernel baseline with pre-set tubular kernel shapes.","marker":"[33]"},{"why":"Defines clDice, the centerline-connectivity metric used to evaluate continuity of tubular structures.","marker":"[40]"},{"why":"SCOPE is the structural-continuity baseline the method is compared against on retinal vessels.","marker":"[53]"},{"why":"U-Net is the backbone used for the layer comparison and for the ablation studies.","marker":"[35]"},{"why":"CHASE DB1 provides the retinal vessel dataset used for the main comparison and ablations.","marker":"[14]"},{"why":"ISBI12 provides the neuron electron-microscopy segmentation benchmark with loops and connections.","marker":"[1]"},{"why":"Supplies the HT29 colon cancer cell dataset from the Broad BioImage Benchmark Collection used as the third evaluation benchmark.","marker":"[3]"}],"fun_headline_variants":["Topology-aware kernels fix connectivity in segmentation","Persistent homology steers convolution to preserve structure","Conformable convolution: topology-guided kernels for accurate segmentation","Adaptive kernels follow topological cues to cut connectivity errors","Persistent homology guides kernel offsets to fix topology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mechanism rests on the assumption that 0-dimensional persistent homology on a single channel-pooled feature map, with a fixed persistence threshold $\\tau_0$, marks the anatomical structures that matter for segmentation at every layer where the module is used.","fun_headline_variants_meta":{"raw":{"variants":["Topology-aware kernels fix connectivity in segmentation","Persistent homology steers convolution to preserve structure","Conformable convolution: topology-guided kernels for accurate segmentation","Adaptive kernels follow topological cues to cut connectivity errors","Persistent homology guides kernel offsets to fix topology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3377,"prompt_tokens":976,"completion_tokens":2401,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2328}},"tokens_in":592,"tokens_out":2401,"duration_ms":16981,"temperature":1.0,"reasoning_tokens":2328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:15:19.440084+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the CHASE vessel model with the TPG prior replaced by Gaussian-dilated random generator locations of the same density; if Betti-zero and Euler errors are no worse than with the true topological prior, the specific topological information is not what drives the improvement. Alternatively, sweep $\\tau_0$ from zero to the 90th percentile of persistence values and check whether the reported error reductions depend sharply on the threshold.","supporting_citations":[{"cited_title":"Scope: Structural continuity preservation for retinal vessel segmen- tation","cited_arxiv_id":null,"evidence_quote":"SCOPE is the structural-continuity baseline the method is compared against on retinal vessels."},{"cited_title":"U- net: Convolutional networks for biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"U-Net is the backbone used for the layer comparison and for the ablation studies."},{"cited_title":"Deformable convolutional networks","cited_arxiv_id":null,"evidence_quote":"Supplies the deformable-convolution mechanism of learnable kernel offsets that Conformable Convolution reuses and redirects."},{"cited_title":"Topological per- sistence and simplification","cited_arxiv_id":null,"evidence_quote":"Supplies the persistence-simplification principle used to filter low-persistence generators in the TPG."},{"cited_title":"Dynamic snake convolution based on topo- logical geometric constraints for tubular structure segmenta- tion","cited_arxiv_id":null,"evidence_quote":"Dynamic Snake Convolution is the main adaptive-kernel baseline with pre-set tubular kernel shapes."},{"cited_title":"cldice-a novel topology-preserving loss function for tubular structure seg- mentation","cited_arxiv_id":null,"evidence_quote":"Defines clDice, the centerline-connectivity metric used to evaluate continuity of tubular structures."},{"cited_title":"An ensemble classification-based approach applied to retinal blood vessel segmentation","cited_arxiv_id":null,"evidence_quote":"CHASE DB1 provides the retinal vessel dataset used for the main comparison and ablations."},{"cited_title":"Crowdsourcing the creation of image segmentation algorithms for connec- tomics","cited_arxiv_id":null,"evidence_quote":"ISBI12 provides the neuron electron-microscopy segmentation benchmark with loops and connections."},{"cited_title":"Cellprofiler: image analysis software for identifying and quantifying cell phenotypes","cited_arxiv_id":null,"evidence_quote":"Supplies the HT29 colon cancer cell dataset from the Broad BioImage Benchmark Collection used as the third evaluation benchmark."}],"review_version":1}