pith. sign in

arxiv: 2606.18723 · v1 · pith:USMFXYBSnew · submitted 2026-06-17 · 💻 cs.CV · cs.LG

Clinically Aligned Geometry Constraints for Robust IVUS Vessel Boundary Segmentation

Yunshu Chen , Litao Yang , Giuseppe Di Giovanni , Jordan Tan , Deval Mehta , Andrew Lin , Derek Chew , Masasi Fujino
show 4 more authors
Julie Butters Stephen Nicholls Zongyuan Ge Kyung Hoon Cho
This is my paper

Pith reviewed 2026-06-26 21:22 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords IVUS segmentationgeometry consistencyvessel boundaryplaque burdendeep learninglumenEEMtopology
0
0 comments X

The pith

GeoCat uses a differentiable geometry loss on diameters and angles to improve IVUS boundary segmentation for clinical use.

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

The paper introduces GeoCat to segment lumen and external elastic membrane boundaries in intravascular ultrasound images by enforcing consistency with clinical geometry measures. Standard segmentation methods often produce boundary errors that distort plaque area and burden calculations, so the authors add a loss term that penalizes inconsistencies in vessel diameters, orientations, and areas during training. The network combines Cartesian and polar views of 5-frame sequences with attention and fusion modules to produce the segmentations. On a dataset of over 12,000 frames from 146 patients, this yields higher overlap scores along with substantially lower errors in the geometric quantities that clinicians use. If the approach works as claimed, it would allow more trustworthy automated quantification of coronary plaque without manual correction.

Core claim

GeoCat is a geometry-consistent network that processes 5-frame IVUS clips using dual Cartesian-polar encoders with cross-domain attention and temporal fusion. It employs a differentiable geometry consistency loss to supervise diameters, orientations, and cross-sectional areas. Trained on 12,242 annotated frames from 146 patients across two IVUS systems, the model achieves a Dice score of 0.93, 95HD of 0.14 mm, 1.0% topology violations, diameter errors of 0.13-0.16 mm, and angular errors of approximately 8 degrees.

What carries the argument

Differentiable geometry consistency loss supervising diameters, orientations, and cross-sectional areas in a dual-encoder network for 5-frame IVUS clips.

If this is right

  • Segmentation reaches Dice of 0.93 and 95HD of 0.14 mm.
  • Topology violation rate falls to 1.0%.
  • Diameter errors drop to 0.13-0.16 mm.
  • Angular errors reach approximately 8 degrees.
  • The outputs support reliable plaque burden quantification.

Where Pith is reading between the lines

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

  • The dual Cartesian-polar encoding could apply to segmentation of other tubular medical structures.
  • Direct clinical-metric supervision might reduce reliance on post-processing steps in analysis software.
  • Results across two commercial systems suggest the constraints help with device variation.
  • Better geometric fidelity could support automated tracking of plaque changes over time.

Load-bearing premise

That supervising diameters, orientations, and areas with a differentiable loss during training will improve clinical geometry metrics on new data without introducing new failure modes.

What would settle it

A held-out test set from a new patient cohort where diameter errors stay above 0.2 mm or topology violations exceed 5 percent.

Figures

Figures reproduced from arXiv: 2606.18723 by Andrew Lin, Derek Chew, Deval Mehta, Giuseppe Di Giovanni, Jordan Tan, Julie Butters, Kyung Hoon Cho, Litao Yang, Masasi Fujino, Stephen Nicholls, Yunshu Chen, Zongyuan Ge.

Figure 1
Figure 1. Figure 1: GeoCat pipeline. (A) Temporal clip sampling. (B) Polar reparameterisation and geometry target derivation. (C) Dual-encoder architecture with temporal attention, cross-domain fusion, and FiLM conditioning. (D) Loss formulation. from 146 patients, GeoCat reduces boundary error by 15% and diameter MAE by 12% relative to the strongest prior method while halving topology violations. 2 Method [PITH_FULL_IMAGE:f… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on three test cases. EEM/lumen contours in red/green; inset tables show per-method clinical errors. GeoCat yields smoother boundaries and lower diameter errors, especially in rows 2–3 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
read the original abstract

Intravascular ultrasound (IVUS) lumen and external elastic membrane (EEM) segmentation is important for quantitative coronary plaque burden assessment. Errors in lumen or EEM delineation directly propagate to plaque area, plaque burden and geometric measurements. However, standard methods prioritising overlap scores often suffer from boundary drift and topology errors, leading to inaccurate clinical measurements. We present GeoCat, a geometry-consistent network that processes 5-frame IVUS clips using dual Cartesian-polar encoders with cross-domain attention and temporal fusion. A differentiable geometry consistency loss directly supervises clinically relevant descriptors including diameters, orientations, and cross-sectional areas. The model is trained on 12,242 annotated frames from 146 patients acquired with two commercial IVUS systems. We evaluate performance using both segmentation accuracy and plaque-relevant clinical metrics, including Dice/IoU, boundary measures(95HD (mm), ASSD), topology violation rate, and clinical geometry errors (dmax/dmin, angles, and areas). On our dataset, GeoCat achieves a Dice of 0.93, reduces 95HD to 0.14 mm, and lowers topology violations to 1.0%. Importantly, it significantly improves geometric fidelity, yielding diameter errors of 0.13-0.16 mm and angular errors of ~8 degrees, supporting reliable plaque burden quantification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents GeoCat, a network for IVUS lumen and EEM segmentation that combines dual Cartesian-polar encoders with cross-domain attention and temporal fusion, plus a differentiable geometry consistency loss supervising diameters, orientations, and cross-sectional areas. Trained on 12,242 frames from 146 patients across two IVUS systems, it reports Dice 0.93, 95HD 0.14 mm, 1% topology violations, and clinical geometry errors of 0.13-0.16 mm (diameters) and ~8° (angles) on held-out data, claiming support for reliable plaque burden quantification.

Significance. If the geometry loss demonstrably improves true clinical fidelity on patient-disjoint held-out data without direct supervision of the reported metrics or undetected failure modes, the work would meaningfully advance IVUS analysis by addressing boundary drift that affects plaque area and burden calculations. The multi-system dataset and joint evaluation on segmentation plus clinical geometry metrics are strengths; however, the absence of ablations, baselines, and loss formulation details limits current impact.

major comments (3)
  1. [Abstract / Methods] Abstract and Methods: the description states that the geometry consistency loss 'directly supervises clinically relevant descriptors including diameters, orientations, and cross-sectional areas,' yet provides no formulation for differentiable extraction of dmax/dmin or area from the mask, nor the loss functional (L1/L2 on scalars). This is load-bearing for the central claim that reported diameter/angular errors reflect improved robustness rather than direct optimization of the evaluation quantities.
  2. [Abstract / Experiments] Abstract and Experiments: no confirmation is given that the train/test split on 12,242 frames from 146 patients is strictly patient-disjoint. Patient-level leakage would undermine the generalization claim that the model supports 'reliable plaque burden quantification' on held-out data.
  3. [Experiments] Experiments: the abstract and reported results contain no baseline comparisons, ablation results on the geometry loss weight, or statistical tests (e.g., paired t-tests on clinical metrics). These omissions directly affect confidence that the 0.13-0.16 mm diameter and ~8° angular improvements are attributable to the proposed loss rather than architecture or data specifics.
minor comments (2)
  1. [Abstract] Notation for clinical descriptors (dmax/dmin, angles) should be defined explicitly when first introduced to avoid ambiguity with standard diameter measurements.
  2. [Results] Figure captions or results tables should clarify whether topology violation rate is computed per frame or per patient to match the clinical metric reporting.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments that identify key areas for clarification and additional analysis. We agree that providing the loss formulation, confirming the data split details, and including baselines/ablations/statistics will strengthen the manuscript. We address each major comment below and will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and Methods: the description states that the geometry consistency loss 'directly supervises clinically relevant descriptors including diameters, orientations, and cross-sectional areas,' yet provides no formulation for differentiable extraction of dmax/dmin or area from the mask, nor the loss functional (L1/L2 on scalars). This is load-bearing for the central claim that reported diameter/angular errors reflect improved robustness rather than direct optimization of the evaluation quantities.

    Authors: We acknowledge that the manuscript omitted the explicit differentiable formulation for extracting dmax/dmin, orientations, and areas from predicted masks, as well as the precise loss functional. This is a valid point that affects reproducibility and the strength of the central claim. In the revised manuscript, we will add a detailed Methods subsection describing the differentiable operations (e.g., contour-based or moment-based extraction) and the loss terms (L1/L2 on the scalars), clarifying how supervision occurs without directly optimizing the held-out evaluation metrics. revision: yes

  2. Referee: [Abstract / Experiments] Abstract and Experiments: no confirmation is given that the train/test split on 12,242 frames from 146 patients is strictly patient-disjoint. Patient-level leakage would undermine the generalization claim that the model supports 'reliable plaque burden quantification' on held-out data.

    Authors: The splits were performed at the patient level with no patient overlap between training and test sets to ensure generalization to unseen patients. We will explicitly state this in the revised Methods and Experiments sections, including the patient counts per split (e.g., details on the 146-patient partitioning) to directly address the generalization claim. revision: yes

  3. Referee: [Experiments] Experiments: the abstract and reported results contain no baseline comparisons, ablation results on the geometry loss weight, or statistical tests (e.g., paired t-tests on clinical metrics). These omissions directly affect confidence that the 0.13-0.16 mm diameter and ~8° angular improvements are attributable to the proposed loss rather than architecture or data specifics.

    Authors: We agree that baselines, loss-weight ablations, and statistical tests would strengthen attribution of the reported gains. In the revision, we will add: comparisons to standard baselines (e.g., Cartesian U-Net and polar variants), ablation experiments varying the geometry loss weight, and paired statistical tests (e.g., t-tests) on the clinical metrics. These will be incorporated into the Experiments section. revision: yes

Circularity Check

0 steps flagged

No significant circularity; geometry loss and reported metrics remain distinct on held-out evaluation

full rationale

The abstract describes a differentiable geometry consistency loss that supervises diameters, orientations, and cross-sectional areas during training on 12,242 frames. Reported results include Dice 0.93, 95HD 0.14 mm, topology violations 1.0%, and clinical errors (diameter 0.13-0.16 mm, angular ~8°). No equations or formulations are given that equate the test-set clinical error quantities directly to the training loss terms by construction. Evaluation uses separate segmentation and clinical metrics on held-out data; the derivation chain does not reduce predictions to inputs via self-definition, fitted-parameter renaming, or self-citation load-bearing steps. The paper is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the modeling choice that a geometry consistency loss will improve clinical metrics; no explicit free parameters, invented entities, or non-standard axioms are stated in the abstract.

pith-pipeline@v0.9.1-grok · 5804 in / 1095 out tokens · 39232 ms · 2026-06-26T21:22:08.631078+00:00 · methodology

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

21 extracted references · 15 canonical work pages

  1. [1]

    Mintz, G.S., Nissen, S.E., Anderson, W.D., et al.: American College of Cardiol- ogy clinical expert consensus document on standards for acquisition, measurement and reporting of IVUS studies.J. Am. Coll. Cardiol.37(5), 1478–1492 (2001). doi:10.1016/S0735-1097(01)01175-5

    work page doi:10.1016/s0735-1097(01)01175-5 2001

  2. [2]

    Balocco, S., Gatta, C., Ciompi, F., et al.: Standardized evaluation methodology and reference database for evaluating IVUS image segmentation.Comput. Med. Imaging Graph.38(2), 70–90 (2014). doi:10.1016/j.compmedimag.2013.07.001

    work page doi:10.1016/j.compmedimag.2013.07.001 2014

  3. [3]

    doi:10.4244/EIJV6I9A195

    Mintz, G.S., Garcia-Garcia, H.M., Nicholls, S.J., et al.: Clinical expert con- sensus document on standards for acquisition, measurement and reporting of IVUS regression/progression studies.EuroIntervention6(9), 1123–1130 (2011). doi:10.4244/EIJV6I9A195

    work page doi:10.4244/eijv6i9a195 2011

  4. [4]

    Yoon, H.J., Hur, S.H.: Optimization of stent deployment by intravascular ultra- sound.Korean J. Intern. Med.27(1), 30–38 (2012). doi:10.3904/kjim.2012.27.1.30

    work page doi:10.3904/kjim.2012.27.1.30 2012

  5. [5]

    doi:10.1016/j.compbiomed.2025.111312

    Zhang,L.,Scholl,B.,Un,S.,etal.:FullyautomatedIVUSimagesegmentationwith efficient deep-learning-assisted annotation.Computers in Biology and Medicine 199, 111312 (2025). doi:10.1016/j.compbiomed.2025.111312

    work page doi:10.1016/j.compbiomed.2025.111312 2025

  6. [6]

    doi:10.1016/j.compmedimag.2024.102336

    Sun, Y., Zhang, X., Gao, G., et al.: Distraction-aware hierarchical learn- ing for vascular structure segmentation in intravascular ultrasound im- ages.Computerized Medical Imaging and Graphics111, 102336 (2024). doi:10.1016/j.compmedimag.2024.102336

    work page doi:10.1016/j.compmedimag.2024.102336 2024

  7. [7]

    doi:10.4070/kcj.2023.0297

    Jeong, J., et al.: Deep learning-based lumen and vessel segmentation of intravas- cular ultrasound images in coronary artery disease.Korean Circulation Journal (2024). doi:10.4070/kcj.2023.0297

    work page doi:10.4070/kcj.2023.0297 2024

  8. [8]

    In: MICCAI 2015, LNCS, vol

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomed- ical image segmentation. In: MICCAI 2015, LNCS, vol. 9351, pp. 234–241. Springer (2015)

    2015

  9. [9]

    In: ECCV 2018, LNCS, vol

    Chen, L.-C., et al.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: ECCV 2018, LNCS, vol. 11211, pp. 833–851. Springer (2018)

    2018

  10. [10]

    doi:10.1186/s12938-023-01171-2

    Dong, L., et al.: Comparison of deep learning-based image segmentation methods for intravascular ultrasound.Biomedical Engineering Online22(1), 111 (2023). doi:10.1186/s12938-023-01171-2

    work page doi:10.1186/s12938-023-01171-2 2023

  11. [11]

    doi:10.1016/j.compmedimag.2021.101939 10 Chen, Yang, Di Giovanni et al

    Szarski, M., Cazottes, A., et al.: Improved real-time segmentation of intravascular ultrasound images using coordinate-aware fully convolu- tional networks.Computerized Medical Imaging and Graphics(2021). doi:10.1016/j.compmedimag.2021.101939 10 Chen, Yang, Di Giovanni et al

    work page doi:10.1016/j.compmedimag.2021.101939 2021

  12. [12]

    doi:10.1016/j.ijcard.2021.06.065

    Gerlach, J., et al.: Advanced deep learning methodology for accurate, real-time seg- mentation of high-resolution intravascular ultrasound images.International Jour- nal of Cardiology338, 169–175 (2021). doi:10.1016/j.ijcard.2021.06.065

    work page doi:10.1016/j.ijcard.2021.06.065 2021

  13. [13]

    doi:10.1016/j.compbiomed.2024.109162

    Bransby, K.M., Bajaj, R., Ramasamy, A., et al.: POLYCORE: Polygon- based contour refinement for improved Intravascular Ultrasound Seg- mentation.Computers in Biology and Medicine182, 109162 (2024). doi:10.1016/j.compbiomed.2024.109162

    work page doi:10.1016/j.compbiomed.2024.109162 2024

  14. [14]

    doi:10.1016/j.media.2023.102922

    Huang, X., Bajaj, R., Li, Y., et al.: POST-IVUS: A perceptual organisation-aware selectivetransformerframeworkforintravascularultrasoundsegmentation.Medical Image Analysis89, 102922 (2023). doi:10.1016/j.media.2023.102922

    work page doi:10.1016/j.media.2023.102922 2023

  15. [15]

    Zhang, J

    Chen, Y., et al.: Geo-UNet: A Geometrically Constrained Neural Framework for Clinical-Grade Lumen Segmentation in Intravascular Ultrasound. In: Xu, X., et al. (eds.) MLMI 2024. LNCS, vol. 15241. Springer, Cham (2025). doi:10.1007/978-3- 031-73284-3_30

    work page doi:10.1007/978-3- 2024

  16. [16]

    doi:10.1148/radiol.2020191145

    Zwanenburg, A., Vallières, M., Abdalah, M.A., et al.: The image biomarker standardization initiative: standardized quantitative radiomics for high- throughput image-based phenotyping.Radiology295(2), 328–338 (2020). doi:10.1148/radiol.2020191145

    work page doi:10.1148/radiol.2020191145 2020

  17. [17]

    In: AAAI, vol

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: FiLM: Visual reasoning with a general conditioning layer. In: AAAI, vol. 32, no. 1, pp. 3942– 3951 (2018)

    2018

  18. [18]

    In: CVPR, pp

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., Xie, S.: A ConvNet for the 2020s. In: CVPR, pp. 11976–11986 (2022)

    2022

  19. [19]

    In: DLMIA/ML-CDS 2018, LNCS, vol

    Zhou, Z., et al.: UNet++: A nested U-Net architecture for medical image segmen- tation. In: DLMIA/ML-CDS 2018, LNCS, vol. 11045, pp. 3–11. Springer (2018)

    2018

  20. [20]

    In: Ad- vances in Neural Information Processing Systems, vol

    Vaswani, A., Shazeer, N., Parmar, N., et al.: Attention is all you need. In: Ad- vances in Neural Information Processing Systems, vol. 30, pp. 5998–6008. Curran Associates (2017)

    2017

  21. [21]

    In: Fourth International Conference on 3D Vision (3DV), pp

    Milletari,F.,Navab,N.,Ahmadi,S.-A.:V-Net:Fullyconvolutionalneuralnetworks for volumetric medical image segmentation. In: Fourth International Conference on 3D Vision (3DV), pp. 565–571. IEEE (2016). doi:10.1109/3DV.2016.79

    work page doi:10.1109/3dv.2016.79 2016