pith. machine review for the scientific record. sign in

arxiv: 2604.17208 · v1 · submitted 2026-04-19 · 💻 cs.CV · cs.AI

Recognition: unknown

CDSA-Net:Collaborative Decoupling of Vascular Structure and Background for High-Fidelity Coronary Digital Subtraction Angiography

Si Li , Chen-Kai Hu , Zhenhuan Lyu , Yuanqing He
Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:19 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords coronary angiographydigital subtraction angiographydeep learningvascular structurebackground restorationboundary artifactsX-ray noise modelingimage decoupling
0
0 comments X

The pith

CDSA-Net decouples vascular structure preservation from background restoration to generate high-fidelity coronary DSA images without boundary artifacts.

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

The paper introduces CDSA-Net to solve motion-related problems that force clinicians to rely on cluttered raw coronary angiograms. Existing deep learning approaches typically leave visible boundary artifacts or distort the native tissue grayscale values needed for confident diagnosis. CDSA-Net explicitly separates the tasks of extracting and preserving vessel structures from restoring a realistic background, then subtracts the restored background from the raw image. A reader would care because the resulting images are intended to support faster morphology and hemodynamic assessments while matching the diagnostic value of unprocessed angiograms. The work therefore targets both technical image quality and practical utility in interventional cardiology.

Core claim

The central claim is that CDSA-Net is the first framework to explicitly decouple and jointly optimize vascular structure preservation and realistic background restoration. It achieves this through a hierarchical geometric prior guidance mechanism inside the coronary structure extraction network that combines integrated geometric priors, gated spatial modulation, and centerline-aware topology loss to maintain vessel continuity. An adaptive noise module inside the coronary background restoration network models the stochastic character of clinical X-ray noise to enable seamless intensity estimation and remove boundary artifacts. The subtracted result is produced directly by subtracting the deno

What carries the argument

Collaborative decoupling framework consisting of CSENet with hierarchical geometric prior guidance for structure preservation and CBResNet with adaptive noise module for background restoration.

If this is right

  • Significantly higher vascular intensity correlation and perceptual quality than prior state-of-the-art methods
  • 25.6 percent improvement in morphology assessment efficiency
  • 42.9 percent gain in hemodynamic evaluation speed
  • Diagnostic consistency with raw angiograms preserved
  • Boundary artifacts eliminated without additional post-processing steps

Where Pith is reading between the lines

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

  • The same decoupling pattern could be tested on motion-affected subtraction tasks in other modalities such as CT perfusion or MR angiography.
  • Embedding the networks in real-time cath-lab software might shorten overall procedure times by reducing the need for manual image corrections.
  • Validation on larger multi-vendor datasets would be required to confirm that the noise modeling generalizes across different X-ray systems.

Load-bearing premise

The adaptive noise module can uniquely model the stochastic nature of clinical X-ray noise to bridge the domain gap, enable seamless background intensity estimation, and achieve complete elimination of boundary artifacts without post-processing.

What would settle it

If clinical angiogram test cases still show visible boundary artifacts after subtraction or yield diagnostic readings that diverge from the raw images, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2604.17208 by Chen-Kai Hu, Si Li, Yuanqing He, Zhenhuan Lyu.

Figure 2
Figure 2. Figure 2: Detailed description of the development of coronary segmentation network model. (a) Development pipeline of the model. (b) Architecture of the proposed network. centerline-aware topology (CAT) loss supervision to enforce structural continuity. This integrated design transformed conventional coronary segmentation into a geometry-constrained reconstruction problem, ensuring that even micro-vessels under seve… view at source ↗
Figure 3
Figure 3. Figure 3: Detailed description of the development of background restoration network model. (a) Development pipeline of the model. (b) Architecture of the proposed network. by coronary territory (left and right coronary angiograms). Synthetic training pairs were generated by superimposing coronary segmentation masks onto pre-contrast backgrounds acquired at the identical projection nomenclature (According to standard… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of segmentation results by different networks. (a) raw angiogram; (b) ground truth; (c) UNet++; (d) TransUNet; (e) Attention U-Net; (f) MedFormer; (g) Proposed CSENet. ections in high-noise regions (shown in rows 1 and 2). Rows 3 and 4 demonstrated that our method extracted vascular structures more accurately in low-contrast areas. Furthermore, rows 5 and 6 illustrated that the proposed method p… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of background restoration results by different networks. (a) synthetic frame; (b) pre-contrast frame; (c) ShadowDiffusion; (d) ShadowFormer; (e) Proposed CBResNet. uced a higher global FID of 41.48 and degraded background metrics (NVPSNR 52.48), indicating hallucinations obscured clinical details. 3. Overall coronary subtraction results 3.1 Qualitative comparison [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of coronary subtraction results. (a) raw angiogram; (b) Zeng’s DeepSA method; (c) Proposed CDSA method. inversely related to the raw angiographic intensities. Therefore, a negative Pearson correlation was expected, and its absolute value reflected the strength of the linear relationship. The proposed method yielded a higher correlation (-0.640 vs. -0.435), indicating stronger preserv… view at source ↗
Figure 7
Figure 7. Figure 7: Exemplary diagram of clinical application for raw and subtracted angiograms by the proposed CDSA method. (a) stenosis detection by raw and subtracted angiograms; (b) TIMI frame count (TFC) assessment by raw and subtracted angiograms. demonstrated that the subtracted data enabled more accurate and rapid TIMI frame count (TFC) assessment due to enhanced contrast. Quantitatively, the method showed high agreem… view at source ↗
read the original abstract

Digital subtraction angiography (DSA) in coronary imaging is fundamentally challenged by physiological motion, forcing reliance on raw angiograms cluttered with anatomical noise. Existing deep learning methods often produced images with two critical clinically unacceptable flaws: persistent boundary artifacts and a loss of native tissue grayscale fidelity that undermined diagnostic confidence. We propose a novel framework termed as CDSA-Net that for the first time explicitly decouples and jointly optimizes vascular structure preservation and realistic background restoration. CDSA-Net introduces two core innovations: (i) A hierarchical geometric prior guidance (HGPG) mechanism, embedded in our coronary structure extraction network (CSENet). It synergistically combines integrated geometric prior (IGP) with gated spatial modulation (GSM) and centerline-aware topology (CAT) loss supervision, ensuring structural continuity. (ii) An adaptive noise module (ANM) within our coronary background restoration network (CBResNet). Unlike standard restoration, ANM uniquely models the stochastic nature of clinical X-ray noise, bridging the domain gap to enable seamless background intensity estimation and the complete elimination of boundary artifacts. The final subtraction is obtained by removing the restored background from the raw angiogram. Quantitatively, it significantly outperformed state-of-the-art methods in vascular intensity correlation and perceptual quality. A 25.6% improvement in morphology assessment efficiency and a 42.9% gain in hemodynamic evaluation speed set a new benchmark for utility in interventional cardiology, while maintaining diagnostic results consistent with raw angiograms. The project code is available at https://github.com/DrThink-ai/CDSA-Net.

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

2 major / 2 minor

Summary. The manuscript introduces CDSA-Net for high-fidelity coronary digital subtraction angiography (DSA). It claims to explicitly decouple vascular structure preservation and background restoration via a hierarchical geometric prior guidance (HGPG) mechanism in CSENet (combining integrated geometric prior, gated spatial modulation, and centerline-aware topology loss) and an adaptive noise module (ANM) in CBResNet. The framework is said to eliminate boundary artifacts without post-processing, preserve native tissue grayscale fidelity, and outperform state-of-the-art methods in vascular intensity correlation and perceptual quality, while delivering 25.6% and 42.9% gains in morphology assessment efficiency and hemodynamic evaluation speed, with diagnostic results consistent with raw angiograms. Code is released publicly.

Significance. If the empirical claims hold after proper validation, this work could advance interventional cardiology by providing artifact-reduced DSA images that improve diagnostic confidence and procedural efficiency over existing deep learning approaches to physiological motion compensation. The explicit joint optimization of structure preservation and stochastic background restoration is a conceptually sound direction, and the public code release at the cited GitHub repository is a clear strength for reproducibility.

major comments (2)
  1. [Abstract] Abstract: The reported quantitative outperformance (vascular intensity correlation, perceptual quality, 25.6% morphology efficiency gain, 42.9% hemodynamic speed gain) is presented without dataset details, baseline comparisons, statistical significance tests, error bars, or ablation studies. This directly affects evaluation of the central claim that the proposed decoupling and ANM yield clinically superior results.
  2. [ANM description (Methods/Architecture)] ANM description (Methods/Architecture): The claim that ANM 'uniquely models the stochastic nature of clinical X-ray noise' to achieve 'complete elimination of boundary artifacts' and 'seamless background intensity estimation' lacks isolating ablations (e.g., vs. Poisson-Gaussian or learned noise priors), quantitative comparison of estimated vs. measured noise statistics on real angiograms, or failure-case analysis. This is load-bearing for attributing performance gains to the collaborative decoupling rather than other architectural factors.
minor comments (2)
  1. [Abstract] The abstract would benefit from explicit definition of the quantitative metrics (e.g., what correlation coefficient or perceptual index is used) and a one-sentence statement of the training/validation dataset sizes and sources.
  2. [Methods] Notation for the joint optimization objective and the precise formulation of the CAT loss could be clarified with an equation or pseudocode block for readers unfamiliar with the subcomponents.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address each major comment point by point below, outlining how we will strengthen the manuscript while maintaining scientific accuracy.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported quantitative outperformance (vascular intensity correlation, perceptual quality, 25.6% morphology efficiency gain, 42.9% hemodynamic speed gain) is presented without dataset details, baseline comparisons, statistical significance tests, error bars, or ablation studies. This directly affects evaluation of the central claim that the proposed decoupling and ANM yield clinically superior results.

    Authors: We agree that the abstract's brevity limits inclusion of supporting details, which are instead provided in the full manuscript (dataset in Section 4.1, baselines and metrics in Table 1 and Section 4.2, statistical significance and error bars in Figures 3-4, ablations in Table 3). To improve accessibility, we will revise the abstract to briefly note the clinical dataset size (e.g., 200 angiograms) and reference the presence of ablations and significance testing in the experiments. This partial update respects abstract constraints while directing readers to the detailed evidence. revision: partial

  2. Referee: [ANM description (Methods/Architecture)] ANM description (Methods/Architecture): The claim that ANM 'uniquely models the stochastic nature of clinical X-ray noise' to achieve 'complete elimination of boundary artifacts' and 'seamless background intensity estimation' lacks isolating ablations (e.g., vs. Poisson-Gaussian or learned noise priors), quantitative comparison of estimated vs. measured noise statistics on real angiograms, or failure-case analysis. This is load-bearing for attributing performance gains to the collaborative decoupling rather than other architectural factors.

    Authors: We acknowledge that stronger isolating evidence would better support attribution of gains to ANM and the decoupling framework. The manuscript currently includes component ablations (Table 4) demonstrating ANM's role in artifact reduction, but lacks direct comparisons to Poisson-Gaussian models or measured noise statistics. In the revision, we will add new experiments: (i) isolating ablations versus Poisson-Gaussian and learned priors, (ii) quantitative noise matching (e.g., variance, histogram comparison) on real angiograms, and (iii) a dedicated failure-case analysis subsection with examples. These changes will clarify the contribution of the proposed approach. revision: yes

Circularity Check

0 steps flagged

No circularity: architecture and empirical claims are independent of inputs

full rationale

The paper introduces CDSA-Net as a new architecture with HGPG (including IGP, GSM, CAT loss) in CSENet and ANM in CBResNet, then reports quantitative gains on vascular correlation, perceptual quality, and clinical efficiency metrics. No equations, fitted parameters, or self-citations are shown that reduce any claimed result to the inputs by construction. The decoupling is presented as an explicit design choice whose value is assessed externally via benchmarks, not derived tautologically. The reader's assessment of score 1.0 aligns with this; the skeptic's concerns address validation gaps rather than circular reasoning.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 2 invented entities

The central claim depends on the effectiveness of newly introduced modules whose performance is asserted without upstream independent evidence or formal derivation in the abstract.

free parameters (1)
  • Network weights and training hyperparameters
    End-to-end trained on unspecified clinical angiogram data; values not reported in abstract.
axioms (2)
  • domain assumption Vascular structures and background can be accurately decoupled using geometric priors and topology supervision without loss of fidelity
    Invoked as the foundation for CSENet and the overall subtraction process.
  • domain assumption Clinical X-ray noise is stochastic and can be adaptively modeled to restore realistic background intensities
    Core premise for the ANM in CBResNet to eliminate boundary artifacts.
invented entities (2)
  • Hierarchical geometric prior guidance (HGPG) mechanism no independent evidence
    purpose: Synergistically combines integrated geometric prior, gated spatial modulation, and centerline-aware topology loss for structural continuity
    Newly proposed component embedded in CSENet.
  • Adaptive noise module (ANM) no independent evidence
    purpose: Models stochastic clinical X-ray noise for background restoration and artifact elimination
    Newly proposed component within CBResNet.

pith-pipeline@v0.9.0 · 5594 in / 1508 out tokens · 47760 ms · 2026-05-10T07:19:52.170483+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

57 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    The protocol was approved by the institutional ethics committee, and written informed consent was obtained

    Patient enrollment and data acquisition This retrospective study included 650 consecutive patients who underwent elective percutaneous coronary intervention (PCI) at the Second Affiliated Hospital of Nanch- ang University (Nanchang, China) between December 2022 and December 2025. The protocol was approved by the institutional ethics committee, and written...

    2022

  2. [2]

    For CSENet, the training set comprised opacified frames from angiograms and their corresponding manually segmented labels

    Proposed coronary digital subtraction angiography (CDSA) framework The dual-network architecture consisted of coronary structure extraction network (CSENet) and a coronary background restoration network (CBResNet). For CSENet, the training set comprised opacified frames from angiograms and their corresponding manually segmented labels. For CBResNet, the t...

  3. [3]

    Coronary structure extraction network (CSENet) development The development of CSENet model primarily involved three core components (Figure 2(a)): (1) generation of manually annotated coronary structure labels, (2) training of the network and (3) validation of the model performance. 3.1 Generation of manually annotated labels Two cardiologists annotated c...

  4. [4]

    on the test set (130 patients), assessing vessel continuity, contour accuracy, and specificity in low -contrast, complex, and indistinct regions. Quantitative evaluation employed Dice Similarity Coefficient (DSC) [22], Centerline Dice Coefficient (clDice) [16], 95% Hausdorff distance (HD95) [23] and Intersection over Union (IoU) [24] against expert annota...

  5. [5]

    4.1 Generation of synthetic training image pairs Here we proposed a dataset synthesis method suitable for the background restoration network

    Coronary background restoration network (CBResNet) development The development of the background intensity restoration network model for coronary angiogram primarily comprised three core components (Figure 3(a)): (1) generation of synthetic training image pairs, (2) training of the restoration network, and (3) validation of the model performance. 4.1 Gene...

  6. [6]

    Overall coronary subtraction The Lambert-Beer law [35] established that X-ray attenuation followed an exponenti- al decay. Let 𝐼0 denoted the incident X -ray intensity, 𝜇𝑏𝑔 and 𝑑𝑏𝑔 represented the attenuation coefficient and thickness of background tissues, and 𝜇𝑐 and 𝑑𝑐 denoted 13 the attenuation coefficient and thickness of the contrast agent. The X-ray...

  7. [7]

    To evaluate the lesion morphology characteristics, experts annotated lesion location (AHA coronary artery segmentation standard [36]) and stenosis severity

    Clinical performance evaluation Three experienced interventional cardiologists evaluated raw angiograms and CDSA- subtracted images in randomized order of 130 consecutive test patients. To evaluate the lesion morphology characteristics, experts annotated lesion location (AHA coronary artery segmentation standard [36]) and stenosis severity. To evaluate th...

  8. [8]

    (a) was raw angiogram, (b) was ground truth, (c) to (f) were UNet++, TransUNet, Attention U -Net, MedFormer, (g) was the proposed CSENet

    Coronary structure extraction network (CSENet) development Figure 4 showed representative results by several methods. (a) was raw angiogram, (b) was ground truth, (c) to (f) were UNet++, TransUNet, Attention U -Net, MedFormer, (g) was the proposed CSENet. Row 2, 4, and 6 were magnified views of rows 1, 3, and 5, respectively. Magenta regions indicated fal...

  9. [9]

    (a) was synthetic frame, (b) was pre-contrast frame, (c) was ShadowDiffusion, (d) was ShadowFormer, (e) was the proposed CBResNet result

    Coronary background restoration network (CBResNet) development Figure 5 showed representative background restoration results. (a) was synthetic frame, (b) was pre-contrast frame, (c) was ShadowDiffusion, (d) was ShadowFormer, (e) was the proposed CBResNet result. Row 2, 4 and 6 were magnified views of rows 1, 3, and 5, respectively. Compared methods showe...

  10. [10]

    Qualitative and quantitative comparation of different background restoration networks

    Overall coronary subtraction results 3.1 Qualitative comparison 19 Table 2. Qualitative and quantitative comparation of different background restoration networks. Analysis type Metric ShadowDiffusion ShadowFormer Proposed Qualitative analysis Overall quality 4.349 4.572 4.722 Restoration fidelity 4.385 4.610 4.759 Transition integrity 4.498 4.684 4.796 Ed...

  11. [11]

    Figure 7(a) was the stenosis detection by raw and subtracted angiograms

    Clinical performance evaluation Figure 7 and Table 3 summarized the clinical performance of the proposed CDSA method. Figure 7(a) was the stenosis detection by raw and subtracted angiograms. The results visually demonstrated that the subtracted data provided clearer visualization comparing with raw angiograms for localizing stenosis and determining its se...

  12. [12]

    The subtraction technology removes non-vascular structures (e.g., vertebral shado- ws, catheter artifacts) while enhancing true vessel signals

    Clinical implications The proposed CDSA technology streamlines catheterization laboratory workflows and accelerates procedural decision-making during time-sensitive interventions where rapid vessel assessment is paramount. The subtraction technology removes non-vascular structures (e.g., vertebral shado- ws, catheter artifacts) while enhancing true vessel...

  13. [13]

    Future work could investi- gate these quantitative functional parameters to fully characterize the technology’s impact on procedural optimization and patient outcomes

    Limitations While demonstrating the efficacy in stenosis evaluation, the clinical validation lacked assessment of advanced applications such as contrast kinetics modeling, hemodynam- 24 ic parameter derivation (e.g., angiography-based FFR [4]). Future work could investi- gate these quantitative functional parameters to fully characterize the technology’s ...

  14. [14]

    Conclusions This study established a deep learning -powered coronary subtraction framework that explicitly decouples and synergistically optimizes vessel structure extraction and bac- kground restoration, embodying the core philosophy of our CDSA -Net architecture. Validated on clinical angiograms, it delivers high-fidelity vessel extraction while eli- mi...

  15. [15]

    All procedures performed were in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki declaration and its later amendments

    Ethic statement This study was approved by the Ethics Committee of the Second Affiliated Hospital of Nanchang University. All procedures performed were in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki declaration and its later amendments. Written informed consent was obtained from all individual p...

    1964

  16. [16]

    Funding This work was supported by Special Innovative Projects of O rdinary Colleges and Universities in Guangdong Province [2024KTSCX139]; Research Start-up Project of Guangdong Institute of Science and Technology [ 2023KYQ182]; Jiangxi Provincial Natural Science Foundation [20232BAB216008]; China Scholarship Council [20250 6820050]. 25

  17. [17]

    Availability of data and materials The datasets generated and/or analy zed during the current study are available in the github repository, [https://github.com/DrThink-ai/CDSA-Net]

  18. [18]

    Competing interests The authors declare that they have no competing interests

  19. [19]

    C -KH: Conceptualization, Data curation, Formal analysis, Resources, Writing-original draft

    Author Contributions SL: Data curation, Investigation, Methodology, Software, Validation, Visualization, Writing-original draft. C -KH: Conceptualization, Data curation, Formal analysis, Resources, Writing-original draft. ZHL: Investigation, Software, Visualization. YQH: Conceptualization, Data curation, Formal analysis, Project administration, Resources,...

  20. [20]

    Blur Invariant Phase Correlation in X -Ray Digital Subtraction Angiography

    Tang SY, Wang YT, Chen YW. Blur Invariant Phase Correlation in X -Ray Digital Subtraction Angiography . IEEE/ICME 2007; Beijing, China, 2007; 1715-1719. https://doi.org/10.1109/ICCME.2007.4382041

    work page doi:10.1109/iccme.2007.4382041 2007

  21. [21]

    Angiographic Characteristics of Intermediate Stenosis of the Left Anterior Descending Artery for Determination of Lesion Significance as Identified by Fractional Flow Reserve

    Biasco L, Pedersen F, Lønborg J, Holmvang L, Helqvist S, Saunam äki K, et al. Angiographic Characteristics of Intermediate Stenosis of the Left Anterior Descending Artery for Determination of Lesion Significance as Identified by Fractional Flow Reserve. Am J Cardiol 2015; 115(11): 1475-1480. https://doi.org/10.1016/j.amjc- ard.2015.02.047

    work page doi:10.1016/j.amjc- 2015

  22. [22]

    Intravascular Imaging -Guided or Angiography-Guided Complex PCI

    Lee JM, Choi KH, Song YB, Lee JY , Lee SJ, Lee SY , et al. Intravascular Imaging -Guided or Angiography-Guided Complex PCI. N Engl J Med 2023; 388(18): 1668-1679. https://doi.org/10.1056/MEJM- oa2216607

    work page doi:10.1056/mejm- 2023

  23. [23]

    Diagnostic and prognostic utility of fractional flow reserve for in-stent restenosis

    AI-Hawwas M, Vallurupalli S, Randhawa S, Hakeem A, Agarwal S, Miler K, et al. Diagnostic and prognostic utility of fractional flow reserve for in-stent restenosis. Cardiovasc Revasc Med 2025; Jun 4: S1553-8389(25)0 0289-1. https://doi.org/10.1016/j.carrev.2025.06.004

    work page doi:10.1016/j.carrev.2025.06.004 2025

  24. [24]

    Using an entropy similarity measure to enhance the quality of DSA images with an algorithm based on template matchi ng

    Buzug TM, Weese J, Fassnacht C, Lorenz C . Using an entropy similarity measure to enhance the quality of DSA images with an algorithm based on template matchi ng. VBC 1996; Lecture Notes in Computer Science,

    1996

  25. [25]

    https://doi.org/10.1007/BFb0046959

    Springer, Berlin, Heidelberg. https://doi.org/10.1007/BFb0046959

    work page doi:10.1007/bfb0046959

  26. [26]

    Image Registration for Digital Subtraction Angiograph y

    Meijering EH, Zuiderveld KJ, Viergever MA . Image Registration for Digital Subtraction Angiograph y. Int J Comput Vis 1999; 31: 227-246. https://doi.org/10.1023/A:1008074100927

    work page doi:10.1023/a:1008074100927 1999

  27. [27]

    An invariant approach for image registration in digital subtraction angiography

    Bentoutou Y , Taleb N, EI Mezouar MC, Taleb M, Jetto L . An invariant approach for image registration in digital subtraction angiography. Pattern Recognit 2002; 35(12): 2853-2865. https://doi.org/10.1016/S0031-320 3(02)00016-X. 26

    work page doi:10.1016/s0031-320 2002

  28. [28]

    A fast image registration algorithm for Digital Subtraction Angiography

    Nejati M, Amirfattahi R and Sadri S. A fast image registration algorithm for Digital Subtraction Angiography. ICBME 2010; Isfahan, Iran, 1-4. https://doi.org/10.1109/ICBME.2010.5704959

    work page doi:10.1109/icbme.2010.5704959 2010

  29. [29]

    Application of ICA to X-ray coronary digital subtraction angiography

    Tang S, Wang Y , Chen Y. Application of ICA to X-ray coronary digital subtraction angiography. NEUROCO- MPUTING 2012; 79: 168-172. https://doi.org/10.1016/j.neucom.2011.10.012

    work page doi:10.1016/j.neucom.2011.10.012 2012

  30. [30]

    Accurate vessel extraction via tensor completion of background layer in X-ray coronary angiograms

    Qin BJ, Jin MX, Hao DD, Lv YS, Liu QG, Zhu YQ , et al. Accurate vessel extraction via tensor completion of background layer in X-ray coronary angiograms. Pattern Recognit 2019; 87: 38-54. https://doi.org/10.101 6/j.patcog.2018.09.015

    2019

  31. [31]

    Deep learning -based digital subtraction angiography images generation

    Gao YF, Song Y , Yin XR, W WW, Zhang L, Chen Y , et al. Deep learning -based digital subtraction angiography images generation. Int J Comput Assist Radiol Surg 2019; 14:1775-1784. https://doi.org/10.100 7/s11548-019-02040-x

    2019

  32. [32]

    Pretrained subtraction and segmentation model for coronary angiograms

    Zeng YJ, Liu H, Hu J, Zhao ZB, She Q . Pretrained subtraction and segmentation model for coronary angiograms. Sci Rep 2024; 14: 19888. https://doi.org/10.1038/s41598-024-71063-5

    work page doi:10.1038/s41598-024-71063-5 2024

  33. [33]

    Background Subtraction Angiog- raphy with Deep Learning Using Multi -frame Spatiotemporal Angiographic Input

    Cantrell DR, Cho L, Zhou CC, Faruqui SHA, Potts MB, Jahromi BS, et al. Background Subtraction Angiog- raphy with Deep Learning Using Multi -frame Spatiotemporal Angiographic Input. Journal of Imaging Informatics in Medicine 2024; 37(134-144). https://doi.org/10.1007/s10278-023-00921-x

    work page doi:10.1007/s10278-023-00921-x 2024

  34. [34]

    Training of a deep learning based digital subtraction angiography method using synthetic data

    Duan LZ, Eulig E, Knaup M, Adamus R, Lell M, Kachelrieb M. Training of a deep learning based digital subtraction angiography method using synthetic data. Medical Physics 2024; 51(4793-4810). https://doi.org/ 10.1002/mp.16973

    work page doi:10.1002/mp.16973 2024

  35. [35]

    Multiscale vessel enhancement filtering

    Frangi AF, Niessen WJ, Vincken KL, Viergever MA. Multiscale vessel enhancement filtering. MICCAI 1998; Lecture Notes in Computer Science, 1496. Springer, Berlin, Heidelberg. https://doi.org/10.1007/BFb005619- 5

    work page doi:10.1007/bfb005619- 1998

  36. [36]

    2021 , url =

    Shit S, Paetzold JC, Sekuboyina A, Ezhov I, Unger A, Zhylka A, et al. clDice -a Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. CVPR 2021; Online, 16555-16564. https://doi.org/10.11- 09/CVPR46437.2021.01629

    work page arXiv 2021

  37. [37]

    ITK-SNAP: An interactive tool for semi-automatic segmentation of multi-modality biomedical images

    Y ushkevich PA, Gao Y, Gerig G. ITK-SNAP: An interactive tool for semi-automatic segmentation of multi-modality biomedical images. EMBC 2016; Orlando, 3342-3345. https://doi.org/10.1109/EMBC.20165 91443

    work page doi:10.1109/embc.20165 2016

  38. [38]

    Hubert’s multi-rater kappa revisited

    Andrés AM, Hernández MÁ. Hubert’s multi-rater kappa revisited. Br J Math Stat Pyschol 2020; 73(1): 1-22. https://doi.org/10.1111/bmsp.12167

    work page doi:10.1111/bmsp.12167 2020

  39. [39]

    A Data-scalable Transformer for Medical Image Segmentation: Architecture , Model Efficiency, and Benchmark [preprint]

    Gao YH, Zhou M, Liu D, Yan ZN, Zhang ST, Metaxas DN. A Data-scalable Transformer for Medical Image Segmentation: Architecture , Model Efficiency, and Benchmark [preprint]. Available via arXiv.org. 2023; https://doi.org/10.48550/arXiv.2203.00131. Accessed 12 Dec 2025

    work page doi:10.48550/arxiv.2203.00131 2023

  40. [40]

    Fcos: Fully convolutional one-stage object detection

    Takikawa T, Acuna D, Jampani V , Fidler S. Gated -SCNN: Gated Shape CNNs for Semantic Segmentation. ICCV2019; Seoul, South Korea, 5229-5238. https://doi.org/10.1109/ICCV .2019.00533

    work page doi:10.1109/iccv 2019

  41. [41]

    A Technique for the Measurement of Attitudes

    Likert R. A Technique for the Measurement of Attitudes. Arch Psychol 1932; 140: 1-55

    1932

  42. [42]

    Measures of the Amount of Ecologic Association Between Species

    Dice LR. Measures of the Amount of Ecologic Association Between Species . Ecol 1945; 26(3): 297-302. https://doi.org/10.2307/1932409

    work page doi:10.2307/1932409 1945

  43. [43]

    Comparing images using the Hausdorff distance

    Huttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing images using the Hausdorff distance. TPAMI 193; 15(9): 850-863. https://doi.org/10.1109/34.232073

    work page doi:10.1109/34.232073

  44. [44]

    UniBox: An Advanced Object Detection Network [preprint]

    Yu JH, Jiang YN, Wang ZY, Cao ZM, Huang T. UniBox: An Advanced Object Detection Network [preprint]. Aailable via arXiv.org.2016; https://doi.org/10.48550.arXiv.1608.01471. Accessed 12 Dec 2025

    work page arXiv 2016

  45. [45]

    UNet++: A Nested U -Net Architecture for Medical Image Segmentation

    Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. UNet++: A Nested U -Net Architecture for Medical Image Segmentation. DLMIA 2018, ML -CDS 2018. L ecture Notes in Computer Science, 11045. Springer, Cham. https://doi.org/10.1007/978-3-303-00889-5_1/. 27

    work page doi:10.1007/978-3-303-00889-5_1/ 2018

  46. [46]

    TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation [preprint]

    Chen JN, Lu YY , Yu QH, Luo XD, Adeli E, Wang Y , et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation [preprint]. Available via arXiv.org.2021; https://doi.org/10.48550/arXiv.210- 2.04306. Accessed 25 Feb 2026

    work page doi:10.48550/arxiv.210- 2021

  47. [47]

    Attention U-Net: Learning Where to Look for the Pancreas

    Oktay O, Schlemper J, Le Fologc L, Lee M, Heinrich M, Misawa K, et al. Attention U -Net: Learning Where to Look for the Pancreas [preprint]. Available via arXiv.org.2018; https://doi.org/10.48550/arXiv.1804.03999. Accessed 25 Feb 2026

    work page internal anchor Pith review doi:10.48550/arxiv.1804.03999 2018

  48. [48]

    ShadowFormer: Global Context Helps Image Shadow Removal [preprint]

    Guo LQ, Huang SY, Liu D, Cheng H, Wen BH. ShadowFormer: Global Context Helps Image Shadow Removal [preprint]. Available via arXiv.org.2023; https://doi.org/10.48550/arXiv.2302.01650. Accessed 12 Dec 2025

    work page doi:10.48550/arxiv.2302.01650 2023

  49. [49]

    2021 , url =

    Hou Q, Zhou D and Feng J. Coordinate Attention for Efficient Mobile Network Design. CVPR 2021; 13708-13717. https://doi.org/10.1109/CVPR46437.2021.01350

    work page doi:10.1109/cvpr46437.2021.01350 2021

  50. [50]

    Statistical Image Reconstruction Using Mixed Poisson-Gaussian Noise Model for X -Ray CT [preprint]

    Ding Q, Long Y , Zhang X, Fessler JA. Statistical Image Reconstruction Using Mixed Poisson-Gaussian Noise Model for X -Ray CT [preprint]. Available via arXiv.org.2018; https://doi.org/10.48550/arXiv.1801.09533. Accessed 27 Feb 2026

    work page doi:10.48550/arxiv.1801.09533 2018

  51. [51]

    Fundamentals of Digital Image processing

    Jain AK. Fundamentals of Digital Image processing. Englewood Cliffs, New Jersey: Prentice Hall; 1989

    1989

  52. [52]

    Image quality assessment: from error visibility to structural similarity

    Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. TIP 2004; 13(4): 600-612. https://doi.org/10.1109/TIP.2003.819861

    work page doi:10.1109/tip.2003.819861 2004

  53. [53]

    SmoothGrad: Removing noise by adding noise.arXiv preprint arXiv:1706.03825, 2017

    Heusel M, Ramsauer H, Unterthiner T, Nessler B, Hochreiter S. GANs Trained by a Two Time -Scale Update Rule Converge to a Local Nash Equilibrium. NeurIPS 2017; 6626-6647. https://doi.org/10.48550/arXiv.1706. 08500

    work page doi:10.48550/arxiv.1706 2017

  54. [54]

    ShadowDiffusion: When Degradation Prior Meets Diffusion Model for Shadow Removal

    Guo LQ, Wang C, Yan WH, Huang SY , Wang YF, Pfister H, et al. ShadowDiffusion: When Degradation Prior Meets Diffusion Model for Shadow Removal. CVPR 2023; 14049-14058. https://doi.org/10.1109/CVPR5272 9.2023.01350

    work page doi:10.1109/cvpr5272 2023

  55. [55]

    The Beer-Lambert Law

    Swinehart DF. The Beer-Lambert Law. J. Chem. Educ. 1962; 39(7): 333. https://doi.org/10.1021/ed039p333

    work page doi:10.1021/ed039p333 1962

  56. [56]

    A reporting system on patients evaluated for coronary artery disease

    Auste WG, Edwards JE, Frye RL, Gensini GG, Gott VL, Griffith LS, et al. A reporting system on patients evaluated for coronary artery disease . Report of the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery, American Heart Assocation. CIRCULATION 1975; 51(4): 5-40. https://doi.org/10.1161/01.cir.51.4.5

    work page doi:10.1161/01.cir.51.4.5 1975

  57. [57]

    TIMI Frame Count: A Quantitative Method of Assessing Coronary Artery Flow

    Gibson CM, Cannon CP, Daley WL, Dodge JT, Alexander B, Marble SJ, et al. TIMI Frame Count: A Quantitative Method of Assessing Coronary Artery Flow. CIRCULATION 1996; 93(5): 879 -888. https://doi.org/10.1161/01.cir.93.5.879

    work page doi:10.1161/01.cir.93.5.879 1996