{"id":"9002925c-55ca-4ca7-8ab6-63362bbf77e2","arxiv_id":"1909.02664","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-case proof-of-concept suggests network metrics such as driver-node fraction and search information can differ between healthy and diseased coronary angiograms.","lead":"The paper turns two coronary angiograms, one healthy and one diseased, into weighted networks of vessel intersections and measures their structure. It argues these network measurements could help automate interpretation of coronary angiography.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that network features distinguish healthy from diseased coronary trees rests on two manually segmented networks from different patients and sources; reported differences are confounded and lack any variance estimate.","rationale":"The strongest claim is that network features derived from coronary angiography can capture differences between healthy and diseased coronary trees and that this 'paves the way' to automated classification (Section 4). For this claim to hold, the derived networks must faithfully represent the coronary trees, and the healthy/diseased contrast must be the only systematic difference between the two networks. Neither condition is met. The networks are hand-built from two images that differ in patient, source, and presumably acquisition parameters; the pixel unit of edge weights alone means magnification differences directly alter routing-efficiency values. The manual tracing is also unquantified: no inter-observer or test-retest variability is reported, and the annotator was not blinded to the labels, so the abundance of Lambda-branches and the driver-node fraction may reflect tracing effort and image quality rather than physiology. This is not merely an absence of error bars; it makes the reported differences unfalsifiable as stated, because the original networks, edge weights, and code are not shared. The reader's verdict was already CONDITIONAL with high confidence, and the paper itself concedes that more than two CAs are needed. I agree that this is a useful methodological proposal worth conditional consideration, but the central empirical contrast is not established. Since the stress-test does not uncover a flaw distinct from the reader's weakest assumption, the verdict should remain unchanged: conditional acceptance pending real cohort validation.","tokens_in":7793,"tokens_out":9424,"duration_ms":99876,"concrete_test":"Run the full pipeline on a cohort of at least 20 healthy and 20 diseased CAs acquired under a single X-ray protocol, using an automated or semi-automated vessel-extraction tool (e.g., centerline extraction or CNN-based segmentation) so that node selection and diameter measurement are not hand-traced. Compute the six Section 3 statistics for every image, blinded to labels, and test separation with cross-validated classification or bootstrapped effect sizes. If the healthy/diseased distributions overlap substantially (e.g., AUC not significantly above 0.5, or between-group effect smaller than the within-protocol spread), the differences reported in Figures 3, 7, 8, and 9 are artifacts of the two chosen images rather than disease-related network structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 states that nodes were identified with graphical filters and the steps were 'manually conducted,' with edge weights computed as diameter times length in pixels. Section 3.1 states the healthy and diseased angiograms 'are not related to each other' and come from different external sources ([27] and [28]). The two networks therefore differ not only in health status but also in patient anatomy, X-ray projection, magnification, contrast, and tracing effort. A human annotator who knew the labels placed every node and measured every vessel. Every headline statistic, including clustering 0.099 vs 0.063, driver nodes 42% vs 37%, the quartile-degree shift, and the search-information maps, is a single value on one network per group, with no error bars, null model, or inter-annotator reproducibility. Because the manual step controls how many small branches are kept, it directly determines degree distribution, Lambda-branch count, leaf count, and driver-node fraction. The authors' own last paragraph concedes 'a rigorous study with more than two CAs should be done to further formalize and validate this approach.' Under these conditions, the claimed between-group differences cannot be attributed to disease, so the Section 4 claim that these features could classify CAs is not supported by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to model the coronary artery tree extracted from coronary angiography (CA) images as a weighted network, in which nodes are vessel intersections and edge weights are diameter x length measured in pixels. Using one healthy and one diseased left-coronary angiogram, the authors compute global network characteristics (Table 1), degree and quartile-degree distributions (Figs. 6-7), integration measures including shortest-path length, routing efficiency, and search information (Fig. 8), and a controllability analysis giving the fraction of driver nodes (Fig. 9). They report that the diseased network has lower clustering, a quartile shift in in-degree, more lambda-branch motifs, larger search information, and fewer driver nodes, and they argue that such network features could eventually be used to automate CA interpretation and classification.","tokens_in":8031,"tokens_out":4104,"duration_ms":47651,"significance":"If validated, the proposed graph abstraction would be a useful supplement to Quantitative Coronary Angiography, adding topologically derived features (degree-distribution shape, routing efficiency, search information, driver-node fraction) to the visual and geometric features currently used. The manuscript applies standard network measures correctly and is transparent about the manual network-construction step and about the need for larger studies. Its main strength is the clarity of the proposed pipeline and the explicit acknowledgement that the current evidence is only a proof-of-concept. However, the reported between-group differences are single values on two manually segmented networks from unrelated patients, so the central claim that these features can capture disease state is not yet supported by the presented evidence.","major_comments":[{"comment":"The entire comparison rests on two manually constructed networks. Section 3 states that nodes were identified with graphical filters and that the steps were 'manually conducted,' and Section 3.1 states that the healthy and diseased angiograms 'are not related to each other' and come from different sources (references [27] and [28]). Consequently, every headline statistic -- average clustering 0.099 vs. 0.063 (Table 1), the quartile-degree shift (Fig. 7), the search-information values (Fig. 8), and the driver-node fractions 42% vs. 37% (Fig. 9) -- is a single value per group with no variance estimate, no blinding of the annotator, and no inter-annotator reproducibility. Because the manual tracing determines how many small branches are retained, it directly controls the degree distribution, the number of leaves, and the driver-node fraction. The observed differences could therefore be artifacts of patient anatomy, X-ray projection, magnification, contrast, or tracing effort rather than disease state. This is load-bearing for the Section 4 claim that the features 'pave the way' to classification; the authors' own final paragraph conceding that 'a rigorous study with more than two CAs should be done' does not cure the present lack of support. A revision must either add a multi-subject, blinded, reproducibility-aware study or explicitly reduce the claim to a methodological proposal.","section":"Section 3 and Section 3.1"},{"comment":"The lambda-branch motif is introduced after inspecting the disease-case angiogram (Figs. 2b and 3) and is then used to explain the quartile-degree shift observed in the same two networks (Section 3.3). This is a confirmatory loop: the motif is defined in hindsight on the same images that are later 'explained' by it. The manuscript does not show that lambda-branch abundance distinguishes the two networks under a degree-preserving null model, nor does it test the proposed neovascularization interpretation (reference [29]) against any independent data. Without such a test, the claimed link between lambda-branch abundance and the pathology is not established.","section":"Section 3.2 and Section 3.3"},{"comment":"The interpretation of driver-node fraction presupposes that 'a healthy network with a high percentage of driver nodes is more resilient to malfunctions.' This axiom is presented as intuitive, but it is not derived from the controllability framework of references [34,35] (which concerns structural controllability of linear dynamics) and is not tested against a model of blood-flow failure in a coronary tree. Moreover, the reported difference (42% vs. 37%) is small, and no sensitivity analysis shows how robust this difference is to plausible perturbations of the manually constructed networks. The resilience claim should either be formally justified or removed; at minimum, the driver-node fraction should be reported with confidence intervals from perturbed network ensembles.","section":"Section 3.5"},{"comment":"The comparison of routing efficiency and search information between the healthy and diseased networks is not quantitatively meaningful in the absence of common calibration. Edge weights are diameter x length in pixels, but the two angiograms come from different patients and different sources, so the average pixel scale and the true vessel diameters are not comparable. The AvgTop10% values (e.g., 18.29 vs. 21.48 for the search-information maps) therefore mix anatomical scale with acquisition scale. The manuscript should normalize edge weights, use dimensionless summary statistics, and report full distributions with variance rather than a single top-10% average.","section":"Section 3.4 and Figure 8"}],"minor_comments":[{"comment":"The second author's affiliation contains typographical errors: 'Devision' should be 'Division' and 'Orthopeadic' should be 'Orthopaedic'.","section":"Author affiliation"},{"comment":"Figure 3 contains stray 'aa' labels in the upper left of both panels; these appear to be leftover annotation artifacts and should be removed.","section":"Figure 3"},{"comment":"The term 'quartile-degree distribution' is not defined. Please specify whether the quartiles are computed over nodes ranked by degree, over degree values, or over edges, and how the bins were chosen.","section":"Figure 7"},{"comment":"Reference [27] ('What is coronary angiography') lacks full bibliographic details, and the source of the healthy angiogram should be described more precisely. Reference [28] should state which figure or panel was used and how the stenosis was confirmed.","section":"References"},{"comment":"The phrase 'no significant difference' in the degree distributions is used without any statistical test. Since the networks are small and fixed, a significance statement is not supported unless a permutation or resampling test is provided.","section":"Section 3.3"},{"comment":"Figure 10 reproduces a figure from Andrikos et al. [9] without an explicit permission statement or a license note; this should be addressed before publication.","section":"Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like an early-stage methods proposal than a validated empirical study. If the journal is open to such proposals, a major revision that adds reproducibility evidence, a null model, and a more cautious framing could make it acceptable. If the journal requires empirical validation for classification claims, consider asking the authors to reframe the contribution as a perspective rather than a demonstration. I found no evidence of problematic citation practice; the self-citation [35] is for a standard algorithm and is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You can read this in fifteen minutes. It is a small proof-of-concept paper. The authors take two coronary angiograms—one labeled healthy, one diseased—manually identify the vessel intersections and measure vessel lengths and diameters in pixels, build weighted networks, and compute standard network statistics on each: degree distributions, clustering, routing efficiency, search information, and driver-node fraction. The finding, in their telling, is that the diseased network has more Λ-branches, a shifted quartile in-degree distribution, slightly worse search information, and fewer driver nodes. The paper is clearly written and the authors are transparent that this is a two-case illustrative study, not a validation. The novelty is modest: they are applying existing network measures to a new imaging object, coronary angiography. That has not been done in the cited literature, and the idea—extracting a quantitative network model from the whole coronary tree rather than measuring individual vessel diameters—is a reasonable direction. The discussion is appropriately cautious and the last paragraph explicitly calls for more cases; I do not read it as a strong empirical claim. The soft spot is the load-bearing one. The two networks are built by hand from images of different patients, different sources, and different acquisition protocols. The annotator knew which case was which. Any measured difference—clustering 0.099 versus 0.063, driver nodes 42% versus 37%, the quartile shift—could reflect annotation effort or imaging details rather than disease. There are no error bars, no inter-annotator check, no null model, and no code or data to reproduce the networks. The paper itself acknowledges this. That means the headline contrasts are anecdotes, not evidence. I agree with the reader's verdict: conditional. The proposal is plausible and worth pursuing, but the reported differences do not discriminate health from disease on the current evidence. I would not cite this for the specific numbers. I would cite it only as an early example of network-based modeling of coronary angiograms if I needed that citation. For peer review, I would send it out, but with a clear expectation: the authors need automated extraction or at least multi-annotator reliability data, a larger cohort, and shared artifacts. As a conference workshop note about a modeling paradigm it is fine. As a demonstration that network features classify coronary disease, it does not yet land.","headline":"A clearly described two-case demonstration of network features on coronary angiograms, but the manual segmentation and missing artifacts make the headline comparisons suggestive rather than established.","tokens_in":8502,"tokens_out":1117,"would_cite":false,"duration_ms":10933,"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":"Modeling a coronary angiogram as a weighted vessel network separates a healthy from a stenosed coronary tree.","keywords":["coronary angiography","complex networks","network controllability","degree distribution","network integration","search information","neovascularization","coronary artery tree"],"falsifier":"Take a cohort of at least 50 left coronary angiograms with independent stenosis labels, build each coronary network with automated segmentation or blinded manual tracing, and recompute the five measures; if the clustering gap, in-degree quartile shift, search-information difference, and driver-node fractions do not separate labeled healthy from diseased images, or if they track the acquisition site instead, the central claim is refuted. The paper's own admitted missing step, a rigorous study with more than two angiograms, is exactly this test.","tokens_in":7588,"feed_emoji":"🫀","tokens_out":8276,"duration_ms":85313,"temperature":0.7,"pith_summary":"This paper proposes that the coronary artery tree visible in an angiogram should be treated as a weighted complex network: vessel intersections become nodes, vessel segments become directed edges, and each edge is weighted by vessel diameter times length. The authors build two such networks by hand, one healthy and one diseased, and show that standard network measurements split them: the diseased network has lower average clustering, more \\Lambda$-branches, an in-degree distribution concentrated in a lower quartile, higher search information, and fewer driver nodes (37% versus 42%). Their argument is that these graph-level differences capture how well blood is supplied, not just how thick individual arteries look, and that network features could therefore feed automated diagnosis. A sympathetic reading takes the paper as a proof-of-concept whose promised payoff is a new class of quantitative features for coronary angiography.","feed_headline":"Coronary arteries as networks reveal disease","feed_subtitle":"Clustering, branch shape, search information, and driver nodes all shift with stenosis in the two networks under study.","key_machinery":"The carrying object is the weighted directed coronary network, built by placing a node at every vessel intersection, an edge along each vessel segment in the direction of blood flow, and a weight equal to the vessel's diameter multiplied by its length in pixels. Around this object the paper deploys five measurement families: global network statistics (node and edge counts, average degree, clustering, diameter), the \\Lambda$-branch motif, degree distributions and their quartile profiles, integration measures (shortest-path length, routing efficiency, search information), and the minimum set of driver nodes from controllability theory. The \\Lambda$-branch is the named motif doing much of the interpretive work, and search information plus driver-node fraction are the two measures the paper singles out as promising classification features.","core_discovery":"The central claim is that the graph structure of the coronary tree carries disease-relevant information that visual inspection and single-vessel quantitative measurements miss. On the two networks constructed here, disease is marked by a 36% lower average clustering coefficient, an abundance of \\Lambda$-branches, an in-degree quartile shifted from the fourth to the third, larger search information between vessel-intersection pairs, and a smaller minimum driver-node set. The paper reads these signatures physiologically: extra \\Lambda$-branches and small emergent vessels indicate neovascularization triggered by insufficient blood supply, and fewer driver nodes mean the diseased network is easier to control but less resilient to localized failure. It concludes that network-based measurements are a needed phase in automating coronary angiography interpretation, with the explicit caveat that validation on more than two angiograms remains to be done.","pith_inferences":["A blinded cohort study with a fixed automated segmentation pipeline is the cleanest test: if the 36% clustering gap, the quartile shift, and the driver-node gap follow the acquisition site rather than the stenosis label, the apparent disease signal is an artifact of the two hand-built networks.","The driver-node result suggests a resilience reading that the paper leaves implicit: healthier coronary networks need more driver nodes and are therefore harder to steer to a desired state but more robust to single-node failure; simulating progressive stenosis-induced branching on a synthetic coronary tree should shrink the minimum driver set.","Search information may grade stenosis severity continuously rather than merely separating healthy from diseased; adding artificial occlusions to one angiogram and recomputing route hiddenness would test whether the measure rises monotonically with occlusion severity.","The same directed, weighted-tree formalism connects coronary networks to other fluid-distribution networks such as leaf venation and river deltas, where \\Lambda$-branch-like motifs and search information are already studied; a cross-system comparison could reveal whether the disease signature is a generic symptom of failing distribution networks."],"forward_implications":["If the network approach is right, clustering, \\Lambda$-branch count, degree quartiles, search information, and driver-node fraction become candidate features for training machine-learning classifiers on coronary angiograms.","The same graph construction can be automated by existing 3D reconstruction and vessel-extraction pipelines, so the method does not depend on manual tracing once such tools are in place.","Network features could sharpen follow-up care, for example by flagging early revascularization after stent implantation before visual changes are obvious.","Because the representation is a tree rather than an imaging modality, the approach extends naturally to non-invasive coronary CT angiography.","The authors themselves state that a rigorous study with more than two angiograms is the necessary next step before these case-study differences can become thresholds."],"supporting_citations":[{"why":"Source of the healthy-case coronary angiogram from which the healthy network is derived.","marker":"[27]"},{"why":"Source of the diseased-case angiogram; the stenosis marked by the green arrow supplies the disease label.","marker":"[28]"},{"why":"Defines the global network characteristics, including degree, clustering, diameter, and shortest paths, used for Table 1 and Figure 8.","marker":"[18]"},{"why":"Underlies the premise that connection patterns can be used to infer functional efficiency of a system.","marker":"[30]"},{"why":"Defines routing (global) efficiency, the cost-weighted measure that brings vessel diameter and length into the integration analysis.","marker":"[31]"},{"why":"Defines search information, the measure of how much information a random walker needs for efficient routing.","marker":"[32]"},{"why":"Provides the network searchability framework behind the search-information comparison in Figure 8.","marker":"[33]"},{"why":"Supplies the algorithm for identifying minimum driver nodes, from which the 42% versus 37% controllability result follows.","marker":"[34]"},{"why":"Supplies the neovascularization mechanism used to explain the abundance of \\Lambda$-branches and small emergent vessels in the diseased network.","marker":"[29]"}],"fun_headline_variants":["Network view of angiograms spots disease","Angiogram networks unmask coronary disease","Graph structure of coronaries shifts with stenosis","Coronary tree network metrics flag illness","Disease visible in coronary network topology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything reported rests on two manually built networks being accurate and comparable: nodes were identified by eye with graphical filters, vessel widths were measured in pixels, and the healthy and diseased images come from different patients and different sources, so any systematic difference in acquisition or tracing could produce the reported network gaps without reflecting disease.","fun_headline_variants_meta":{"raw":{"variants":["Network view of angiograms spots disease","Angiogram networks unmask coronary disease","Graph structure of coronaries shifts with stenosis","Coronary tree network metrics flag illness","Disease visible in coronary network topology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000285,"raw_usage":{"total_tokens":1650,"prompt_tokens":885,"completion_tokens":765,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":701}},"tokens_in":501,"tokens_out":765,"duration_ms":8224,"temperature":1.0,"reasoning_tokens":701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:43:34.164194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a cohort of at least 50 left coronary angiograms with independent stenosis labels, build each coronary network with automated segmentation or blinded manual tracing, and recompute the five measures; if the clustering gap, in-degree quartile shift, search-information difference, and driver-node fractions do not separate labeled healthy from diseased images, or if they track the acquisition site instead, the central claim is refuted. The paper's own admitted missing step, a rigorous study with more than two angiograms, is exactly this test.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the healthy-case coronary angiogram from which the healthy network is derived."},{"cited_title":"Fatal subacute stent thrombosis induced by guidewire fracture with retained filaments in the coronary artery","cited_arxiv_id":null,"evidence_quote":"Source of the diseased-case angiogram; the stenosis marked by the green arrow supplies the disease label."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the global network characteristics, including degree, clustering, diameter, and shortest paths, used for Table 1 and Figure 8."},{"cited_title":"Efficient behavior of small-world networks","cited_arxiv_id":null,"evidence_quote":"Underlies the premise that connection patterns can be used to infer functional efficiency of a system."},{"cited_title":"van den Heuvel, Richard F","cited_arxiv_id":null,"evidence_quote":"Defines routing (global) efficiency, the cost-weighted measure that brings vessel diameter and length into the integration analysis."},{"cited_title":"Hide-and-seek on complex networks","cited_arxiv_id":null,"evidence_quote":"Defines search information, the measure of how much information a random walker needs for efficient routing."},{"cited_title":"Rosvall, A","cited_arxiv_id":null,"evidence_quote":"Provides the network searchability framework behind the search-information comparison in Figure 8."},{"cited_title":"Controllability of complex networks","cited_arxiv_id":null,"evidence_quote":"Supplies the algorithm for identifying minimum driver nodes, from which the 42% versus 37% controllability result follows."},{"cited_title":"Moreno, K-Raman Purushothaman, Marc Sirol, Andrew P","cited_arxiv_id":null,"evidence_quote":"Supplies the neovascularization mechanism used to explain the abundance of \\Lambda$-branches and small emergent vessels in the diseased network."}],"review_version":1}