Pith. sign in

REVIEW 4 major objections 5 minor 55 references

This paper establishes CARDIAG, a multi-center benchmark dataset for pixel-level SYNTAX classification of coronary angiograms, and reports that a ConvNeXt V2 encoder paired with a DeepLabV3+ decoder is currently the strongest single model (

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:39 UTC pith:J62VYN6A

load-bearing objection A genuinely useful dataset and broad architecture sweep, but the 'best model' nomination doesn't survive the error bars; the value is in the resource, not the ranking. the 4 major comments →

arxiv 2607.22139 v1 pith:J62VYN6A submitted 2026-07-24 cs.CV cs.AIcs.LG

CARDIAG: A Dense Segment Classification Benchmark of Deep Learning Architectures for Coronary Angiography

Dominik Bernard Lau , Hubert Malinowski , Jerzy Szyjut , Adam Brzeski , Tomasz Dziubich , Rados{\l}aw Targo\'nski , Tomasz Figatowski , Natalia Zieli\'nska This is my paper
classification cs.CV cs.AIcs.LG
keywords SYNTAX segmentationcoronary angiographysemantic segmentationbenchmark datasetdeep learning architecturesmodel calibrationuncertainty masksensembling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper establishes a standardized benchmark for the task of labeling every pixel of a coronary X-ray angiogram with one of the 26 SYNTAX coronary artery segment classes (plus background and catheter). The authors introduce CARDIAG, a multi-center dataset of 644 annotated frames with explicit train/validation/test splits that prevent patient-level leakage, and evaluate 24 deep learning architectures under identical training conditions. They report that a modern convolutional network pairing the ConvNeXt V2 encoder with the DeepLabV3+ decoder is the strongest single model (macro F1 = 0.456), and that ensembling it with a Mamba-based U-Net and a Feature Pyramid Network pushes performance to 0.479. The paper also shows that models are well calibrated, that leave-one-center-out generalization is stable, that data volume rather than architecture is the main bottleneck, and that near-zero F1 scores on distal segments likely reflect an annotation ceiling rather than pure algorithmic failure. A sympathetic reader would care because the field previously lacked a reproducible, leakage-controlled benchmark, and this dataset plus protocol give future SYNTAX segmentation and lesion-detection work a common measuring stick.

Core claim

On the paper's own terms, the central discovery is a ranking: for dense SYNTAX segment classification of coronary angiograms, the best single architecture is a ConvNeXt V2 encoder with a DeepLabV3+ decoder, reaching a macro F1 of 0.456 across the non-background classes, with the best ensemble (ConvNeXt V2 + Mamba U-Net + FPN, soft/hard voting) reaching 0.479. The authors argue this shows that translation-equivariant convolutions with multi-scale atrous spatial pyramid pooling are better suited to vascular structures than Vision Transformers or state-space models alone, while Mamba models excel specifically at diameter fidelity. They further claim that all tested architectures are well calibr

What carries the argument

The load-bearing component is the encoder–decoder pairing of ConvNeXt V2 with a DeepLabV3+ head: the encoder provides translation-equivariant hierarchical features, while the decoder's atrous spatial pyramid pooling (ASPP) samples those features at multiple dilation rates and fuses them with a low-level skip, letting the model keep both coarse semantic context and fine vessel borders. The ensemble combines this model with a Mamba U-Net (a state-space sequence model that captures long-range artery context) and a Feature Pyramid Network (which merges multi-scale feature maps), using three voting schemes. The CARDIAG dataset itself is a second kind of machinery: its patient-level, center-strati

Load-bearing premise

The ranking rests on the assumption that the pixel-level ground truth, produced by three specialists from one institution with inter-observer agreement measured on only 10 images (DSC 0.87, 0.907 after uncertainty masks, Fleiss κ=0.761 for stenosis), is accurate enough to discriminate between models that are within a few points of each other.

What would settle it

Re-annotate the test set's distal and small segments (e.g., 16a, 16b, 16c, 9a, 10) by a consensus panel and compute F1 for those segments alone; if consensus labels yield materially higher scores for any architecture than the current near-zero values, the annotation-ceiling explanation for distal failures is wrong. Alternatively, if a single architecture other than ConvNeXt V2 + DeepLabV3+ — for instance a Vision Transformer trained on the full CARDIAG data plus ARCADE — exceeds macro F1 0.456 on the same test split, the paper's nomination of the best single architecture is overturned.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • SYNTAX segmentation now has a public benchmark with fixed patient-stratified splits, so future architectures can be compared on the same footing instead of on in-house datasets.
  • Ensembling a modern CNN with a state-space model and a feature pyramid network improves macro F1 from 0.456 to 0.479, so complementary inductive biases are a cheap route to better vessel labeling.
  • Because Mamba U-Net has the lowest diameter error (DE=2.176), state-space models are the most promising basis for downstream stenosis severity measurement, even though they do not top the F1 ranking.
  • Data volume, not architecture choice, is the binding constraint: the best model at 50% training data matches most other models at 100%, and the trend is still rising.
  • Near-zero F1 on distal segments (16a, 16b, 16c, 9a, 10, 10a, 14, 15) should be read as an annotation ceiling; improving labels, not just models, is needed to make progress there.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: If the annotation-ceiling hypothesis for distal segments is correct, the meaningful headroom for the benchmark lies in the clinically critical and high-importance segments (F^ci_1 ≈ 0.69, F^hi_1 ≈ 0.46), and research attention should shift to those.
  • Inference: The authors' leave-one-center-out results show a large drop on Center 2, whose test set is much larger than its training set; this suggests distribution shift across centers is real and that reporting per-center rather than pooled numbers will be necessary for clinical deployment claims.
  • Inference: Because the dataset ships uncertainty masks and supplementary frames, a natural next experiment is to train an aleatoric-uncertainty-aware model that outright ignores the masked distal regions during loss computation; if that raises F1 on remaining segments, the annotation-ceiling claim gains support.
  • Inference: A useful test of the author's ranking claim: retrain the top three architectures on a substantially larger multi-center dataset (e.g., by adding ARCADE) and check whether ConvNeXt V2 + DeepLabV3+ still tops a transformer such as SegFormer Big; the paper's own data-efficiency curves suggest the ordering could change with scale.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces CARDIAG, a multi-center coronary angiography dataset with dense SYNTAX segment labels, uncertainty masks, catheter masks, and DICOM metadata (644 samples from 114 patients across five centers). It benchmarks 24 segmentation architectures, ranging from CNNs to ViTs and state-space models, under a patient- and center-stratified split, and evaluates them with a battery of metrics (macro F1, segment-importance F1, diameter error, clDice, HD95, calibration error). The authors nominate ConvNeXt V2 + DeepLabV3+ as the best single architecture (macro F1 = 0.456) and report that an ensemble with Mamba U-Net and FPN increases macro F1 to 0.479. They further report leave-one-center-out generalization, data-efficiency curves, calibration, and patient/acquisition meta-analyses. The dataset, model weights, and code are released.

Significance. If the claims are supported, the paper would make a valuable contribution: it addresses the scarcity of public SYNTAX segmentation benchmarks, takes careful steps against patient- and center-level leakage, provides a broad architecture comparison, and releases the data and code as community resources. The multi-metric evaluation (F1 variants, diameter error, topology-aware clDice, Hausdorff distance) is a strength, as is the inclusion of uncertainty masks and metadata. However, the central ranking claim and the calibration claim currently outrun the statistical evidence; the paper itself flags key limitations (no inter-annotation divergence analysis, uncertainty about the calibration metric) that must be resolved or explicitly qualified before the benchmark can serve as the 'standardized evaluation protocol' the abstract promises.

major comments (4)
  1. [§3.1, Table 2] The central claim that ConvNeXt V2 + DeepLabV3+ is 'best performing' is not statistically supported. The table reports ± intervals but never defines what they are; if they are across test images, the top model (0.456±0.027) overlaps with VMUnet (0.438±0.023), FPN (0.437±0.025), DeepLabV3+ (0.432±0.024), Swin-UMamba (0.427±0.024), and Mamba-UNet (0.427±0.023). No significance tests are reported, no correction is made for comparing 24 models, and the winner is selected on the same test set used for ranking. The observed differences may be selection noise. Please define the intervals, add per-image and per-class uncertainty estimates, and report paired significance tests (e.g., bootstrap or Wilcoxon) with multiple-comparison control, or explicitly reframe the conclusion as identifying a top group rather than a single best architecture.
  2. [§2.4, §3.1 calibration claim] The abstract and Section 3.1 claim that 'all the architectures are well calibrated,' but the only evidence is C = AUC ROC(U, E) using five dropout samples. This metric measures whether uncertainty correlates with errors; it is not a calibration measure (agreement between predicted probabilities and empirical frequencies). The text itself admits 'we are not fully convinced of the metric C being the best indicator of that.' This is an internal contradiction. Provide a proper calibration metric (e.g., expected calibration error or reliability diagrams) or remove/qualify the calibration claim. The current claim is load-bearing for the stated contribution of 'Uncertainty Quantification for Clinical Safety.'
  3. [§2.1 and Limitations paragraph] The paper acknowledges in Limitations that 'we have not analyzed inter-annotation divergence therefore we do not know about aleatoric uncertainty which might ultimately be a glass ceiling of any possible segmentation methods.' Yet the benchmark uses annotations from three specialists at one institution, with inter-observer validation on only 10 images and no per-segment agreement analysis. Section 3.1 attributes near-zero F1 scores for distal segments (16a,16b,16c,9a,10,10a,14,15) to a possible 'intrinsic labeling ceiling.' Since macro F1 is averaged over all 26 classes, annotation noise in these distal classes can directly affect the ranking. Please report per-segment inter-observer agreement or, at minimum, state explicitly that rankings on distal segments are within annotation uncertainty and may not reflect algorithmic quality.
  4. [§3.2, Table 3] The claim that 'the overall findings confirm that the trained models maintain strong generalization capabilities' is not supported by the presented evidence. Table 3 (labeled 'Qualitative Results' despite containing numeric scores) shows Center 2 with macro F1 around 0.38, substantially below Center 1 (0.54), and the 'Average' row appears to be an unweighted mean of the five centers rather than a patient-level pooled estimate. The Center 2 condition is also qualitatively different: the training set (279 images) is smaller than the test set (366 images). No statistical comparison across centers is provided. Please report pooled estimates with patient-level bootstrap intervals, and discuss the Center 2 result as a possible distribution shift rather than dismissing it solely due to test set size.
minor comments (5)
  1. [§2.4, Eq. (8)] The definition C = AUC ROC(U, E) is incomplete: the text does not specify whether U and E are computed per pixel, per image, or per segment, nor how the five dropout samples are aggregated. Please clarify.
  2. [Table 2] The column header 'CHD 95' is inconsistent with the Hausdorff distance HD95 defined in Eq. (5). Define the notation or correct the header. Also, the '±' values are not defined in the table caption or text.
  3. [§2.2 and Table 1] The paper refers to 'ConvNeXt V2 encoder with DeepLab V3 Plus decoder' in the abstract/Introduction, but Table 1 lists 'ConvNeXtV2 DeepLab' and other naming variants. Use one consistent naming convention throughout.
  4. [Figure 8 caption] The caption says the ensemble consists of 'VM-UNet, ConvNeXt V2 Deeplab and FPN,' while Section 2.2 and the abstract state the ensemble is ConvNeXt V2 + FPN + Mamba U-Net. Please correct the inconsistency.
  5. [Discussion, Limitations] There are typos: 'aleatoircuncertainty' and 'hisegments' should be 'aleatoric uncertainty' and 'these segments.' Also, the female age range is reported as 69–91 while the overall range is 34–87; if the female range is correct, the overall range needs clarification.

Circularity Check

0 steps flagged

No significant circularity: the benchmark's central claims are empirical measurements, not derivations that reduce to their inputs.

full rationale

The paper's central claim, nominating ConvNeXt V2 + DeepLabV3+ as the best single architecture (macro F1 = 0.456) and reporting an ensemble improvement to 0.479, is an empirical result obtained by training 24 architectures on a fixed train/validation/test split and reading metrics from the test set. There is no fitted parameter renamed as a prediction: no equation defines the benchmark outcome in terms of itself, and the reported F1, clDice, DE, and HD95 values are measured on held-out images. The only self-referential element is that the authors built both the CARDIAG dataset and the benchmark, and appear to have selected the ensemble members after observing benchmark results; that is test-set-based model selection, a validity concern but not definitional circularity. The Limitations section's admission that inter-annotation divergence was not analyzed is an honest statement of missing evidence, not a circular step. No load-bearing argument reduces to a self-citation; the broad comparison against independent architecture baselines provides external content. The absence of significance testing and undefined ± intervals weaken the ranking's statistical support, but that is a correctness/robustness limitation, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The benchmark has no free parameters in the derivation sense; the load-bearing choices are the label schema, ground-truth reliability, and the fixed global hyperparameter configuration. The main risk is that the reported ranking is conditional on these choices and on the unvalidated uncertainty metric.

free parameters (1)
  • Global training hyperparameter set (AdamW lr, weight decay, EMA, early stopping, loss weights) = lr=1e-4, λ=1e-2, β=(0.5,0.999), T0=20, patience=100, max_epochs=700, L=CE+Dice
    §2.3 states these were chosen empirically via pilot experiments and fixed for all architectures; no per-model tuning or ablation is reported, so the architecture ranking is conditional on this one configuration.
axioms (5)
  • domain assumption SYNTAX score segment definitions [2] are an appropriate pixel-level label taxonomy for coronary angiography.
    Introduced in §2.1; all 26 classes and the F1-metric depend on this schema. If the taxonomy is ambiguous at distal segments, benchmark scores inherit that ambiguity.
  • domain assumption Expert manual labels are reliable ground truth at pixel level despite limited validation.
    Only 10 images were inter-observer validated (DSC 0.87, after uncertainty masks 0.907; Fleiss κ=0.761 for stenosis), §2.1. Distal segments with near-zero F1 may be label-limited rather than model-limited, acknowledged in §3.1.
  • domain assumption Diameter error from distance transform (Eq. 7) measures clinically meaningful vessel width.
    DE is used to rank models and support clinical claims, but no phantom or clinical validation of diameter accuracy is provided.
  • domain assumption Inference-time dropout AUROC (Eq. 8) is a valid proxy for calibration.
    Used in §3.1 to claim all models are well calibrated; the authors themselves doubt the metric, and no ECE or reliability diagrams are given.
  • domain assumption Leave-one-center-out results generalize to unseen hospitals.
    Table 3 shows large variance and imbalanced center sizes; no external validation beyond the five included centers is performed.

pith-pipeline@v1.3.0-alltime-deepseek · 18715 in / 13449 out tokens · 131598 ms · 2026-08-01T05:39:28.886655+00:00 · methodology

0 comments
read the original abstract

Accurate pixel-level classification of coronary angiograms is critical for cardiovascular disease assessment, yet the field lacks standardized evaluation protocols. In this work we demonstrate a new benchmark for the assessment of deep learning models which densely classify pixels of coronary angiograms to one of SYNTAX classes (or background). The evaluation covers 24 distinct architectures starting with classic convnets to recent state-space-based vision algorithms. We release CARDIAG - a multi-center, multi-label dataset which we carefully split to reliably compute metrics, accounting for diameter error, overlap, centerline quality and calibration. The data contains SYNTAX labels, binary, uncertainty and segmentation masks as well as intermediate frames together with the selected non-sensitive DICOM metadata. From the multitude of algorithms, we nominate ConvNeXt V2 encoder with DeepLab V3 Plus decoder as the best performing, achieving macro $F_1=0.456$, which we then ensemble with Mamba U-Net and Feature Pyramid Network, for an increased $F_1=0.479$. We demonstrate all the architectures to be well calibrated and determine the generalization of the top 5 methods, together with the data efficiency of these architectures. We highlight the importance of both high-resolution and low-resolution features in encoding. We also demonstrate the model correctness in the context of patient demographic, vessel sides and projection angle configurations. Overall the released benchmark allows for future studies to robustly and rigorously assess the proposals, not only for SYNTAX segmentation, but lesion detection and many more.

Figures

Figures reproduced from arXiv: 2607.22139 by Adam Brzeski, Dominik Bernard Lau, Hubert Malinowski, Jerzy Szyjut, Natalia Zieli\'nska, Rados{\l}aw Targo\'nski, Tomasz Dziubich, Tomasz Figatowski.

Figure 1
Figure 1. Figure 1: Example comparison of annotations provided by two experts: (a) depicts a full segmentation of the vascular tree, while (b) illustrates segment-specific annotations (RCA mid). In the comparison visualizations, the first image shows the mask generated by ex￾pert 1, the second shows the mask generated by expert 2, the third highlights the differences between annotations (white indicates overlap￾ping regions, … view at source ↗
Figure 2
Figure 2. Figure 2: Segment statistics including left-to￾right ratio, frequencies and average sizes [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Patient demographics statistics of the dataset. m4a, where m4a(x, y) = 1 ⇐⇒ (x, y) ∈ segment 4a, logically m4a(x, y) = 0 otherwise. (4) catheter binary mask: analogously to the labels before, we provide a mask point￾ing to where the catheter used to inject con￾trast dye is located on the image. (5) sup￾plementary frames: we attach to each sam￾ple extra frames for each projection including background backgr… view at source ↗
Figure 4
Figure 4. Figure 4: Labelled X-ray angiography images overlaid on (α, β)-coordinate system [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Summary of the labels available for a single sample human population is not reflected in the data and thus the models using the dataset as base should only be utilized for the aforementioned age gap. This is visualized in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Various model metrics w.r.t. GFLOPs, mind the log scale of X axis [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Architectures performance across SYNTAX segments with highlighted ensem￾bles. In [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of inference results across five selected examples. The dis [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: F1 across various acquisition per￾spectives. Bubble size denotes cluster size. 3.4 Acquisition and patient-specific parameters [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Impact of training data volume on the predictive accuracy of the evaluated models [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: F1 across segments belonging to dif￾ferent sides. For LCA the performance is sig￾nificantly worse, as expected, due to its entan￾glement [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Patient demographic influencing re￾sults. by the distribution of samples and specific cases along the population. An intuitive observation is that the performance is worse for the left side than for the right (see [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 14 canonical work pages · 4 internal anchors

  1. [1]

    The application of deep learning for the segmentation and classifi- cation of coronary arteries,

    S. Kaba, H. Haci, A. Isin, A. Ilhan, and C. Conkbayir, “The application of deep learning for the segmentation and classifi- cation of coronary arteries,”Diagnostics, vol. 13, no. 13, p. 2274, Jul. 2023, doi: 10.3390/diagnostics13132274

    work page doi:10.3390/diagnostics13132274 2023

  2. [2]

    The SYNTAX score: An angiographic tool grading the complexity of coronary artery disease,

    G. Sianoset al., “The SYNTAX score: An angiographic tool grading the complexity of coronary artery disease,”EuroInterven- tion, vol. 1, no. 2, pp. 219–227, Aug. 2005

    2005

  3. [3]

    Progressive perception learning for main coronary segmentation in X-Ray angiography,

    H. Zhang, Z. Gao, D. Zhang, W. K. Hau, and H. Zhang, “Progressive perception learning for main coronary segmentation in X-Ray angiography,”IEEE Trans. Med. Imaging, vol. 42, no. 3, pp. 864–879, Mar. 2023, doi: 10.1109/TMI.2022.3219126

    arXiv 2023

  4. [4]

    Selective ensemble methods for deep learning segmentation of major vessels in invasive coronary angiography,

    J. Parket al., “Selective ensemble methods for deep learning segmentation of major vessels in invasive coronary angiography,” 14 Med. Phys., vol. 50, no. 12, pp. 7822–7839, Jun. 2023, doi: 10.1002/mp.16554

    doi:10.1002/mp.16554 2023

  5. [5]

    T-Net: Nested encoder–decoder architecture for the main vessel segmen- tation in coronary angiography,

    T. J. Jun, J. Kweon, Y.-H. Kim, and D. Kim, “T-Net: Nested encoder–decoder architecture for the main vessel segmen- tation in coronary angiography,”Neural Netw., vol. 128, pp. 216–233, Aug. 2020, doi: 10.1016/j.neunet.2020.05.002

    doi:10.1016/j.neunet.2020.05.002 2020

  6. [6]

    Main coronary vessel segmentation using deep learning in smart medical,

    Z. Xian, X. Wang, S. Yan, D. Yang, J. Chen, and C. Peng, “Main coronary vessel segmentation using deep learning in smart medical,”Math. Probl. Eng., vol. 2020, pp. 1–9, Oct. 2020, doi: 10.1155/2020/8858344

    work page doi:10.1155/2020/8858344 2020

  7. [7]

    Deep learning model for coronary artery segmentation and quan- titative stenosis detection in angiographic images,

    B. Huanget al., “Deep learning model for coronary artery segmentation and quan- titative stenosis detection in angiographic images,”Med. Phys., vol. 52, no. 7, Jul. 2025, doi: 10.1002/mp.17970

    doi:10.1002/mp.17970 2025

  8. [8]

    Coronary angiography im- age segmentation based on PSPNet,

    X. Zhu, Z. Cheng, S. Wang, X. Chen, and G. Lu, “Coronary angiography im- age segmentation based on PSPNet,” Comput. Methods Programs Biomed., vol. 200, p. 105897, Mar. 2021, doi: 10.1016/j.cmpb.2020.105897

    arXiv 2021

  9. [9]

    G2ViT: Graph neu- ral network-guided vision transformer en- hanced network for retinal vessel and coro- nary angiograph segmentation,

    H. Xu and Y. Wu, “G2ViT: Graph neu- ral network-guided vision transformer en- hanced network for retinal vessel and coro- nary angiograph segmentation,”Neural Netw., vol. 176, p. 106356, Aug. 2024, doi: 10.1016/j.neunet.2024.106356

    arXiv 2024

  10. [10]

    Optimizing ensemble U-Net architectures for robust coronary vessel segmentation in angiographic im- ages,

    S.-S. Changet al., “Optimizing ensemble U-Net architectures for robust coronary vessel segmentation in angiographic im- ages,”Sci. Rep., vol. 14, no. 1, Mar. 2024, doi: 10.1038/s41598-024-57198-5

    doi:10.1038/s41598-024-57198-5 2024

  11. [11]

    HR-UMamba++: A high- resolution multi-directional mamba frame- work for coronary artery segmentation in X-Ray coronary angiography,

    X. Zhang, P. Lu, Z. Zheng, and W. Li, “HR-UMamba++: A high- resolution multi-directional mamba frame- work for coronary artery segmentation in X-Ray coronary angiography,”Fractal Fract., vol. 10, no. 1, p. 43, Jan. 2026, doi: 10.3390/fractalfract10010043

    work page doi:10.3390/fractalfract10010043 2026

  12. [12]

    Dataset for automatic region-based coronary artery disease diag- nostics using X-Ray angiography images,

    M. Popovet al., “Dataset for automatic region-based coronary artery disease diag- nostics using X-Ray angiography images,” Sci. Data, vol. 11, no. 1, Jan. 2024, doi: 10.1038/s41597-023-02871-z

    work page doi:10.1038/s41597-023-02871-z 2024

  13. [13]

    Multivessel Coronary Artery Segmentation and Stenosis Localisation using Ensemble Learning

    M. Bilalet al., “Multivessel coronary artery segmentation and stenosis local- isation using ensemble learning,”arXiv preprint arXiv:2310.17954, 2023, doi: 10.48550/ARXIV.2310.17954

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.17954 2023

  14. [14]

    MPSeg : Multi-Phase strategy for coronary artery Segmentation

    J. Ku, Y.-H. Lee, J. Shin, I. K. Lee, and H.-W. Kim, “MPSeg : Multi-phase strategy for coronary artery segmenta- tion,”arXiv preprint arXiv:2311.10306, 2023, doi: 10.48550/ARXIV.2311.10306

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.10306 2023

  15. [15]

    YOLO-Angio: An Algorithm for Coronary Anatomy Segmentation

    T. Liu, H. Lin, A. K. Katsaggelos, and A. Kline, “YOLO-Angio: An algo- rithm for coronary anatomy segmenta- tion,”arXiv preprint arXiv:2310.15898, 2023, doi: 10.48550/ARXIV.2310.15898

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.15898 2023

  16. [16]

    Accurate segmentation and la- beling of coronary artery segments in X- Ray angiography with an improved UNet- based cGAN architecture,

    Q. Yang, H. Yi, L. Yi, M. Liu, and X. Chen, “Accurate segmentation and la- beling of coronary artery segments in X- Ray angiography with an improved UNet- based cGAN architecture,”Biomed. Signal Process. Control, vol. 112, p. 108812, Feb. 2026, doi: 10.1016/j.bspc.2025.108812

    arXiv 2026

  17. [17]

    UENet: A novel generative adversarial network for angiography image segmentation,

    X. Shi, T. Du, S. Chen, H. Zhang, C. Guan, and B. Xu, “UENet: A novel generative adversarial network for angiography image segmentation,” in Proc. 42nd Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Montreal, QC, Canada, Jul. 2020, pp. 1612–1615, doi: 10.1109/EMBC44109.2020.9175334

    arXiv 2020

  18. [18]

    Training and validation of a deep learning architecture for the auto- matic analysis of coronary angiography,

    T. Duet al., “Training and validation of a deep learning architecture for the auto- matic analysis of coronary angiography,” EuroIntervention, vol.17, no.1, pp.32–40, May 2021, doi: 10.4244/eij-d-20-00570

    work page doi:10.4244/eij-d-20-00570 2021

  19. [19]

    Coronary artery vascular segmentation on limited data via pseudo-precise la- bel,

    M. Zhai, T. Du, R. Yang, and H. Zhang, “Coronary artery vascular segmentation on limited data via pseudo-precise la- bel,” inProc. 41st Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), Berlin, Germany, Jul. 2019, pp. 816–819, doi: 10.1109/EMBC.2019.8856682

    arXiv 2019

  20. [20]

    AGMN: Association graph-based graph matching network for 15 coronary artery semantic labeling on inva- sive coronary angiograms,

    C. Zhaoet al., “AGMN: Association graph-based graph matching network for 15 coronary artery semantic labeling on inva- sive coronary angiograms,”Pattern Recog- nit., vol. 143, p. 109789, Nov. 2023, doi: 10.1016/j.patcog.2023.109789

    arXiv 2023

  21. [21]

    EAGMN: Coronary artery se- mantic labeling using edge attention graph matching network,

    C. Zhao, Z. Xu, G.-U. Hung, and W. Zhou, “EAGMN: Coronary artery se- mantic labeling using edge attention graph matching network,”Comput. Biol. Med., vol. 166, p. 107469, Nov. 2023, doi: 10.1016/j.compbiomed.2023.107469

    arXiv 2023

  22. [22]

    Multi-graphgraphmatchingfor coronary artery semantic labeling in inva- sive coronary angiograms,

    C. Zhao, Z. Xu, P. Baral, M. Esposito, and W.Zhou, “Multi-graphgraphmatchingfor coronary artery semantic labeling in inva- sive coronary angiograms,”Pattern Recog- nit., vol. 169, p. 111943, Jan. 2026, doi: 10.1016/j.patcog.2025.111943

    arXiv 2026

  23. [23]

    HAGMN-UQ: Hyper associa- tion graph matching network with uncer- tainty quantification for coronary artery semantic labeling,

    C. Zhao, M. Esposito, Z. Xu, and W. Zhou, “HAGMN-UQ: Hyper associa- tion graph matching network with uncer- tainty quantification for coronary artery semantic labeling,”Med. Image Anal., vol. 99, p. 103374, Jan. 2025, doi: 10.1016/j.media.2024.103374

    arXiv 2025

  24. [24]

    Diffusion-based user-guided data augmentation for coronary stenosis detection,

    S. Seo, I. K. Lee, H.-W. Kim, J. Min, and C.-H. Jung, “Diffusion-based user-guided data augmentation for coronary stenosis detection,” inProc. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), Dae- jon, RepublicofKorea, Sep.2025, pp.149– 169

    2025

  25. [25]

    Self-supervised ves- sel segmentation via adversarial learn- ing,

    Y. Maet al., “Self-supervised ves- sel segmentation via adversarial learn- ing,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Montreal, QC, Canada, Oct. 2021, pp. 7516–7525, doi: 10.1109/ICCV48922.2021.00744

    arXiv 2021

  26. [26]

    Auto- matic segmentation of coronary arteries in X-Ray angiograms using multiscale anal- ysis and artificial neural networks,

    F. Cervantes-Sanchez, I. Cruz-Aceves, A. Hernandez-Aguirre, M. A. Hernandez- Gonzalez, and S. E. Solorio-Meza, “Auto- matic segmentation of coronary arteries in X-Ray angiograms using multiscale anal- ysis and artificial neural networks,”Appl. Sci., vol. 9, no. 24, p. 5507, Dec. 2019, doi: 10.3390/app9245507

    work page doi:10.3390/app9245507 2019

  27. [27]

    CoronaryDominance: Angiogram dataset for coronary domi- nance classification,

    I. Kruzhilovet al., “CoronaryDominance: Angiogram dataset for coronary domi- nance classification,”Sci. Data, vol. 12, no. 1, Feb. 2025, doi: 10.1038/s41597-025- 04676-8

    doi:10.1038/s41597-025- 2025

  28. [28]

    CardioSyntax: End-to-end SYNTAX score prediction - dataset, benchmark and method,

    A. Ponomarchuket al., “CardioSyntax: End-to-end SYNTAX score prediction - dataset, benchmark and method,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (W ACV), Tucson, AZ, USA, Feb. 2025, pp. 5873–5883, doi: 10.1109/WACV61041.2025.00573

    arXiv 2025

  29. [29]

    Digital imag- ing and communications in medicine (DICOM),

    National Electrical Manufacturers Association (NEMA), “Digital imag- ing and communications in medicine (DICOM),” NEMA, Rosslyn, VA, USA, Tech. Rep., 2026, standard PS3 / ISO 12052. [Online]. Available: https://dicom.nema.org/medical/dicom/2026b/

    2026

  30. [30]

    2026 heart disease and stroke statistics: A re- port of US and global data from the American Heart Association,

    L. P. Palaniappanet al., “2026 heart disease and stroke statistics: A re- port of US and global data from the American Heart Association,”Circula- tion, vol. 153, no. 9, Mar. 2026, doi: 10.1161/cir.0000000000001412

    doi:10.1161/cir.0000000000001412 2026

  31. [31]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” inProc. Med. Image Comput. Comput.-Assist. In- terv. (MICCAI), Munich, Germany, 2015, pp. 234–241, doi: 10.1007/978-3-319- 24574-4_28

    doi:10.1007/978-3-319- 2015

  32. [32]

    Attention U-Net: Learn- ing where to look for the pancreas,

    O. Oktayet al., “Attention U-Net: Learn- ing where to look for the pancreas,” in Proc. Med. Image Deep Learn. (MIDL), Amsterdam, The Netherlands, 2018

    2018

  33. [33]

    3D MRI brain tu- mor segmentation using autoencoder regularization,

    A. Myronenko, “3D MRI brain tu- mor segmentation using autoencoder regularization,” inProc. Med. Image Comput. Comput.-Assist. Interv. (MIC- CAI) Brainlesion Workshop (BrainLes), Granada, Spain, 2019, pp. 311–320, doi: 10.1007/978-3-030-11726-9_28

    doi:10.1007/978-3-030-11726-9_28 2019

  34. [34]

    nnU- Net: A self-configuring method for deep learning-based biomedical image segmen- tation,

    F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU- Net: A self-configuring method for deep learning-based biomedical image segmen- tation,”Nat. Methods, vol. 18, no. 2, pp. 203–211, Dec. 2020, doi: 10.1038/s41592- 020-01008-z. 16

    doi:10.1038/s41592- 2020

  35. [35]

    Pyramid scene parsing network,

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” inProc. IEEE Conf. Comput. Vis. Pat- tern Recognit. (CVPR), Honolulu, HI, USA, Jul. 2017, pp. 6230–6239, doi: 10.1109/CVPR.2017.660

    doi:10.1109/cvpr.2017.660 2017

  36. [36]

    Fea- ture pyramid networks for object de- tection,

    T.-Y. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Fea- ture pyramid networks for object de- tection,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Honolulu, HI, USA, Jul. 2017, pp. 936–944, doi: 10.1109/CVPR.2017.106

    doi:10.1109/cvpr.2017.106 2017

  37. [37]

    Unified perceptual parsing for scene understanding,

    T. Xiao, Y. Liu, B. Zhou, Y. Jiang, and J. Sun, “Unified perceptual parsing for scene understanding,” inProc. Eur. Conf. Comput. Vis. (ECCV), Munich, Germany, 2018, pp. 432–448, doi: 10.1007/978-3- 030-01228-1_26

    doi:10.1007/978-3- 2018

  38. [38]

    Rethinking atrous convo- lution for semantic image segmentation,

    L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convo- lution for semantic image segmentation,” arXiv preprint arXiv:1706.05587, 2017, doi: 10.48550/ARXIV.1706.05587

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.05587 2017

  39. [39]

    Encoder- decoder with atrous separable convolution forsemanticimagesegmentation,

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder- decoder with atrous separable convolution forsemanticimagesegmentation,” inProc. Eur. Conf. Comput. Vis. (ECCV), Mu- nich, Germany, 2018, pp. 833–851, doi: 10.1007/978-3-030-01234-2_49

    doi:10.1007/978-3-030-01234-2_49 2018

  40. [40]

    Geometric deep learn- ing: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veličković, “Geometric deep learn- ing: Grids, groups, graphs, geodesics, and gauges,”arXiv preprint arXiv:2104.13478, 2021, doi: 10.48550/ARXIV.2104.13478

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.13478 2021

  41. [41]

    SegFormer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “SegFormer: Simple and efficient design for semantic segmentation with transformers,” inAdv. Neural Inform. Process. Syst. (NeurIPS), 2021, pp. 12077–12090

    2021

  42. [42]

    SwinUNETR-V2: Stronger swin transformers with stagewise convolutions for 3D medical image seg- mentation,

    Y. He, V. Nath, D. Yang, Y. Tang, A. My- ronenko, and D. Xu, “SwinUNETR-V2: Stronger swin transformers with stagewise convolutions for 3D medical image seg- mentation,” inProc. Med. Image Comput. Comput.-Assist. Interv. (MICCAI), Van- couver, BC, Canada, 2023, pp. 416–426, doi: 10.1007/978-3-031-43901-8_40

    work page doi:10.1007/978-3-031-43901-8_40 2023

  43. [43]

    ConvNeXt V2: Co- designing and scaling convnets with masked autoencoders,

    S. Wooet al., “ConvNeXt V2: Co- designing and scaling convnets with masked autoencoders,” inProc. IEEE/CVF Conf. Comput. Vis. Pat- tern Recognit. (CVPR), Vancouver, BC, Canada, Jun. 2023, pp. 16133–16142

    2023

  44. [44]

    SegNeXt: Rethinking convolutional attention design for semantic segmentation,

    M.-H. Guo, C.-Z. Lu, Q. Hou, Z.-N. Liu, M.-M. Cheng, and S.-M. Hu, “SegNeXt: Rethinking convolutional attention design for semantic segmentation,” inAdv. Neu- ral Inform. Process. Syst. (NeurIPS), New Orleans, LA, USA, 2022

    2022

  45. [45]

    MedNeXt: Transformer- driven scaling of convnets for medical im- age segmentation,

    S. Royet al., “MedNeXt: Transformer- driven scaling of convnets for medical im- age segmentation,” inProc. Med. Image Comput. Comput.-Assist. Interv. (MIC- CAI), Vancouver, BC, Canada, 2023, pp. 405–415, doi: 10.1007/978-3-031-43901- 8_39

    work page doi:10.1007/978-3-031-43901- 2023

  46. [46]

    Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs

    X. Ding, X. Zhang, Y. Zhou, J. Han, G. Ding, and J. Sun, “Scaling up your kernels to 31x31: Revisiting large kernel design in CNNs,”arXiv preprint arXiv:2203.06717, 2022, doi: 10.48550/ARXIV.2203.06717

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2203.06717 2022

  47. [47]

    Mamba-UNet: UNet-like pure visual mamba for medical image segmen- tation,

    Z. Wang, J.-Q. Zheng, Y. Zhang, G. Cui, and L. Li, “Mamba-UNet: UNet-like pure visual mamba for medical image segmen- tation,”arXiv preprint arXiv:2402.05079, 2024, doi: 10.48550/ARXIV.2402.05079

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.05079 2024

  48. [48]

    SegMamba: Long-range sequen- tial modeling mamba for 3D medical im- age segmentation,

    Z. Xing, T. Ye, Y. Yang, G. Liu, and L. Zhu, “SegMamba: Long-range sequen- tial modeling mamba for 3D medical im- age segmentation,” inProc. Med. Image Comput. Comput.-Assist. Interv. (MIC- CAI),Marrakesh, Morocco, 2024, pp.578– 588, doi: 10.1007/978-3-031-72111-3_54

    work page doi:10.1007/978-3-031-72111-3_54 2024

  49. [49]

    TorchXRayVision: A library of chest X-Ray datasets and mod- els,

    J. P. Cohenet al., “TorchXRayVision: A library of chest X-Ray datasets and mod- els,” inProc. Med. Image Deep Learn. (MIDL), Zurich, Switzerland, 2022

    2022

  50. [50]

    RadImageNet: An open radiologic deep learning research dataset 17 for effective transfer learning,

    X. Meiet al., “RadImageNet: An open radiologic deep learning research dataset 17 for effective transfer learning,”Radiol. Ar- tif. Intell., vol. 4, no. 5, Sep. 2022, doi: 10.1148/ryai.210315

    doi:10.1148/ryai.210315 2022

  51. [51]

    Averaging weights leads to wider optima and better generaliza- tion,

    P. Izmailovet al., “Averaging weights leads to wider optima and better generaliza- tion,” inProc. Conf. Uncertainty Artif. In- tell. (UAI), Monterey, CA, USA, 2018

    2018

  52. [52]

    Metrics for evaluating 3D medical image segmenta- tion: Analysis, selection, and tool,

    A. A. Taha and A. Hanbury, “Metrics for evaluating 3D medical image segmenta- tion: Analysis, selection, and tool,”BMC Med. Imaging, vol. 15, no. 1, Aug. 2015, doi: 10.1186/s12880-015-0068-x

    doi:10.1186/s12880-015-0068-x 2015

  53. [53]

    Comparing images using the Hausdorff distance,

    D. Huttenlocher, G. Klanderman, and W. Rucklidge, “Comparing images using the Hausdorff distance,”IEEE Trans. Pat- tern Anal. Mach. Intell., vol. 15, no. 9, pp. 850–863, 1993, doi: 10.1109/34.232073

    doi:10.1109/34.232073 1993

  54. [54]

    clDice - a novel topology- preserving loss function for tubular struc- ture segmentation,

    S. Shitet al., “clDice - a novel topology- preserving loss function for tubular struc- ture segmentation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2021, pp. 16555–16564, doi: 10.1109/CVPR46437.2021.01629

    arXiv 2021

  55. [55]

    Lumen diameter of nor- mal human coronary arteries. influence of age, sex, anatomic variation, and left ven- tricular hypertrophy or dilation

    J. T. Dodge, B. G. Brown, E. L. Bolson, and H. T. Dodge, “Lumen diameter of nor- mal human coronary arteries. influence of age, sex, anatomic variation, and left ven- tricular hypertrophy or dilation.”Circula- tion, vol. 86, no. 1, pp. 232–246, Jul. 1992, doi: 10.1161/01.cir.86.1.232. 18 Table 2: Benchmark quantitative results. ModelF (m) 1 F (ci) 1 F (hi...

    work page doi:10.1161/01.cir.86.1.232 1992