Comparative Analysis of Deep Learning Models for Coronary Artery Segmentation: Performance and Inference Time Evaluation

Comparative Analysis of Deep Learning Models for Coronary Artery Segmentation: Performance and Inference Time Evaluation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Comparative Analysis of Deep Learning Models for Coronary Artery Segmentation: Performance and Inference Time Evaluation Benny Uhoranishema, Mediatrice Dusenge, Yollanda Umutoni, Theophilla Igihozo, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7467606/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Background Coronary artery disease (CAD) is one of the leading causes of cardiovascular mortality worldwide. Accurate segmentation of coronary arteries from X-ray coronary angiography (XCA) images is crucial for assessing vessel morphology and stenosis, thereby supporting computer-aided diagnosis and guiding interventional treatment decisions. Although recent studies have primarily focused on enhancing segmentation accuracy using deep learning models, limited attention has been given to evaluating their inference time — a factor that is equally important for clinical deployment and real-time decision support. Objective This study compares both segmentation performance and inference time of U-Net, U-Net++, and SegFormer on the ARCADE XCA dataset (stenosis and SYNTAX subsets). Methods All XCA images were resized to 256 × 256 pixels, normalized, and augmented prior to training. The U-Net and U-Net + + architectures were implemented as convolutional encoder–decoder networks with skip connections, whereas SegFormer employed a hierarchical Transformer-based encoder coupled with a lightweight MLP decoder. All models were trained for 100 epoch using cross-entropy loss with class-balancing weights. Performance was evaluated in terms of segmentation accuracy, dice score, and per-image inference time. Results On the stenosis subset, U-Net and U-Net + + achieved the highest training accuracy (99.82%), while SegFormer attained a slightly lower accuracy (99.15%) but delivered the fastest inference time (0.05 s per image). On the SYNTAX subset, U-Net + + obtained the best training accuracy (98.13%), followed closely by U-Net (98.04%) and SegFormer (97.00%). Despite its lower accuracy, SegFormer consistently demonstrated superior efficiency, achieving the shortest inference time (0.18 s per image). Conclusion U-Net + + demonstrated the highest segmentation accuracy, SegFormer provided the most significant runtime advantage, and U-Net achieved a balanced trade-off between the two. Taken together, these findings suggest that model selection should be informed by the specific priorities of clinical deployment, whether accuracy, inference speed, or a compromise between both is most critical. Coronary artery disease Coronary angiography Deep learning Image segmentation Inference time Figures Figure 1 Figure 2 Figure 3 1. Background Coronary artery disease (CAD) is one of the leading causes of cardiovascular mortality worldwide, contributing substantially to the global burden of cardiovascular diseases ( 1 – 4 ). Accurate detection and localization of coronary artery stenosis are therefore crucial for guiding effective clinical decision-making. Coronary angiography continues to serve as the gold standard for diagnosing CAD; however, manual interpretation of angiograms is time-consuming, subject to observer variability, and prone to diagnostic errors ( 5 – 11 ). The scale of this challenge is underscored by global and regional statistics. Cardiovascular diseases (CVDs) remain the leading cause of death worldwide, contributing to a substantial number of fatalities and disabilities. In 2021, the World Health Organization (WHO) reported that CVDs were responsible for 20.5 million deaths, representing nearly one-third of all global deaths. Although once regarded as conditions of affluence, CVDs now disproportionately affect low- and middle-income countries (LMICs), where more than three-quarters of related deaths occur ( 12 ). On the other hand, CAD is the most common type of heart disease caused by plaque buildup in the walls of the arteries that supply blood to the heart ( 13 ). Therefore, this progressive accumulation of plaque leads to coronary artery stenosis, a narrowing of the arterial lumen that restricts blood flow and can trigger angina, myocardial infarction, and other serious cardiovascular events. Despite these challenges in cardiovascular diseases particularly CAD, Deep Learning (DL) has emerged as a highly promising approach for coronary artery segmentation. It offers precise, reproducible, and automated solutions for both coronary artery segmentation and stenosis localization, thereby holding great potential to transform traditional clinical practice. In 2021, researchers in Ref. ( 14 – 21 ) made significant progress in applying deep learning to coronary artery segmentation and analysis. Existing architectures such as U-Net and Dense U-Net were widely adopted, while several new models were introduced, including multiresolution and multiscale U-Net, cascaded V-shaped networks, DeepCap (capsule networks), and DeepDiscern (dual-network system). These approaches achieved notable improvements in segmentation accuracy, robustness against class imbalance, and computational efficiency, with Dice scores reaching up to 0.97, thereby demonstrating strong potential for coronary artery assessment. Next, from 2022 to 2024, studies consistently advanced coronary artery segmentation by refining established models such as U-Net, 3D U-Net, and DenseNet, while also introducing a few new architectures including Di-VNet, DBCU-Net, PlaqueNet, and GhostNet-U-Net ( 22 – 37 ). These efforts steadily improved segmentation accuracy, efficiency, and clinical applicability across modalities such as CCTA (Coronary Computed Tomography Angiography), XCA, and quantitative angiography. More importantly, while drafting this paper in mid-2025, Mahendiran et al. published AngioPy Segmentation: An open-source, user-guided deep learning tool for coronary artery segmentation, which integrates user-defined ground-truth points to improve segmentation accuracy and reduce the need for manual correction. The model achieved an average F1 score of 0.927 on internal datasets and 0.924 on external validation, demonstrating excellent agreement with established quantitative coronary angiography (QCA) systems and highlighting its potential to enhance clinical accuracy and efficiency ( 38 ). Besides, Abedin et al. introduced DenseSelfU-Net and DenseSelfMA-Net , two Self-Organizing Neural Network (Self-ONN) enhanced frameworks designed for coronary artery segmentation and stenosis localization. On the ARCADE challenge dataset, DenseSelfU-Net achieved a Dice score of 90.35% for full artery segmentation, while DenseSelfMA-Net configurations reached up to 60.92% Dice for stenosis localization, outperforming prior methods such as StenU-Net and SSASS ( 39 ). Starting in 2022, researchers began publishing comprehensive and systematic reviews on deep learning applications for coronary artery imaging. These reviews, including meta-analyses, have primarily examined CCTA and angiography, systematically assessing deep learning applications for coronary artery segmentation, stenosis detection, plaque quantification, and risk prediction. Collectively, they highlight the high diagnostic accuracy of DL models, their growing clinical potential, and the need for multicenter validation and standardized methodologies ( 40 – 44 ). One study by Wang et al. proposed a 3D–2D convolutional network for coronary artery segmentation in angiographic videos, which not only improved vessel detail preservation compared to U-Net and CE-Net, but also achieved inference times of 0.054–0.061 seconds per frame, thereby meeting the requirements for real-time clinical use ( 45 ). While numerous studies have focused primarily on improving segmentation accuracy, only a very limited number have explicitly evaluated inference time. To the best of our knowledge, this is the first study to provide a comparative analysis of deep learning models for coronary artery segmentation by jointly assessing both performance and inference efficiency, an aspect that is essential for real-time clinical deployment. 2. Methodology The proposed system employs a DL–based pipeline for coronary artery segmentation from computed tomography angiography (CTA) images. The workflow comprises three main stages: data preprocessing, model training, and performance evaluation. Preprocessing includes image enhancement techniques to improve vessel visibility and facilitate more accurate feature learning. Deep learning models—specifically U-Net, U-Net++, and SegFormer—were trained to perform pixel-level segmentation of coronary arteries, enabling precise extraction of vessel structures critical for cardiovascular disease assessment. 3.1. Dataset This study utilized the publicly available Automatic Coronary Artery Disease Evaluation (ARCADE) dataset ( 46 , 47 ). The dataset is organized into two main objectives: (i) vessel branch classification according to the SYNTAX methodology (syntax folder), and (ii) stenosis detection (stenosis folder). Each objective contains three subsets (train, val, and test), with corresponding images in “.png” format extracted from DICOM recordings, and annotation files provided in “.JSON” format. The annotation files include image metadata (ID, width, height, file name), categorical labels (26 SYNTAX-based vessel regions or stenosis classes), and detailed mask information including segmentation points, bounding box coordinates, and area. The dataset was carefully annotated by medical experts to support research on automated risk assessment and diagnosis of coronary artery disease ( 46 , 47 ). 3.2. Segmentation Models 3.2.1. U-Net The U-Net architecture follows a symmetric encoder–decoder structure with skip connections, enabling precise localization in biomedical image segmentation tasks. The contracting path consists of repeated applications of two 3×3 convolutions (unpadded), each followed by a rectified linear unit (ReLU), and a 2×2 max-pooling operation with stride 2 for down sampling, doubling the number of feature channels at each step. The expansive path mirrors this process, where each step involves sampling of the feature map, followed by a 2×2 up-convolution that halves the number of feature channels, concatenation with the corresponding cropped feature map from the contracting path, and two successive 3×3 convolutions with ReLU activation. Cropping is applied to compensate for border pixel loss during convolution. At the final layer, a 1×1 convolution maps the 64-component feature vectors to the desired number of classes. In total, the network comprises 23 convolutional layers and exhibits the characteristic U-shaped topology ( 48 ). 3.2.2. UNet++ U-Net + + was proposed by Zhou et al. as an extension of the original U-Net for medical image segmentation. It modifies the baseline U-Net by introducing a dense network of skip connections between the encoder and decoder paths. The architecture links the encoder and decoder through a series of nested, dense skip pathways, which progressively refine feature maps before fusion. Unlike U-Net, where encoder features are directly transmitted to the decoder, U-Net + + maps encoder outputs to the decoder via dense convolutional blocks within redesigned skip pathways. This design reduces the semantic gap between encoder and decoder feature maps, ensuring that the feature representations at corresponding levels are more consistent. As a result, U-Net + + achieves better alignment between low-level spatial details and high-level semantic information, thereby improving segmentation performance ( 49 ). 3.2.3. SegFormer SegFormer is a simple yet efficient semantic segmentation framework that combines a hierarchical Transformer encoder with a lightweight multilayer perceptron (MLP) decoder. The encoder produces multistage features without relying on positional encoding, which improves robustness when the testing resolution differs from training. To further enhance efficiency, SegFormer reshapes and projects the key matrix in self-attention, significantly reducing computational cost. The decoder employs a Mix-FFN block, which integrates depth-wise 3×3 convolution into the MLP to efficiently inject positional information while keeping the parameter count low. By avoiding complex decoders, SegFormer achieves an excellent balance between accuracy and efficiency in semantic segmentation ( 50 ). 3.3. Evaluation Metrics For model evaluation, we employed standard segmentation metrics ( 51 , 51 ). Accuracy measures the fraction of correctly classified pixels and is defined as: $$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$ 1 where TP represents true positives, TN true negatives, FP false positives, and FN false negatives. Accuracy is straightforward to interpret and suitable for balanced datasets; however, it may be misleading for imbalanced data, as it does not account for the relative impact of false positives and false negatives. $$\:Dice\:Score=\frac{2TP}{2TP+FP+FN}$$ 2 In terms of TP, FP, FN: Dice score referred to as the F1-score in machine learning, ranges from 0 to 1 and is particularly effective for image segmentation and object detection tasks, as it emphasizes the correct identification of positive cases. More importantly, it is widely used in medical image segmentation because it emphasizes the correct identification of positive (foreground) pixels and provides a more reliable measure of segmentation quality than accuracy alone. On the other hand, Dice measures the overlap between the predicted segmentation and the ground truth, where X represents the set of predicted positive pixels and Y represents the set of ground truth positive pixels. $$\:Dice\:Score=\frac{2\:.\:\:X\cap\:Y\:}{\left|X\right|+\left|Y\right|}$$ 3 3. Results and Discussion This section reports the segmentation performance of U-Net, U-Net++, and SegFormer on the stenosis and SYNTAX datasets. Performance was assessed using training and validation accuracy, per-image Dice similarity coefficients (DSC), and inference time. The Dice score, widely used in medical image segmentation, quantifies the overlap between predicted masks and ground truth annotations for each image. A value of 1 indicates perfect overlap, whereas 0 indicates no overlap. Reporting per-image Dice scores allows us to capture variability across cases and better understand model generalization. Moreover, all model training, validation, prediction, and inference time assessments were conducted on an NVIDIA GeForce RTX 4070 GPU. 3.1. Results on Stenosis Dataset 3.1.1. Segmentation Performance Analysis All three models achieved very high segmentation accuracy on the stenosis dataset. U-Net and U-Net + + both reached validation accuracies of 99.82%, while SegFormer slightly lagged at 99.15%. Representative training dynamics of U-Net++ (the best performer on both subsets) are shown in Fig. 1 , demonstrating stable convergence with minimal overfitting. Table 1 presents the per-image Dice scores on a representative set of stenosis test images, along with the average across the dataset. U-Net + + consistently produced the highest Dice values, reflecting superior overlap with the ground truth. U-Net followed closely, while SegFormer showed slightly lower scores, particularly in cases with complex vessel boundaries. Beyond the numerical results, qualitative analysis further confirmed the quantitative findings and offered a more intuitive understanding of model behavior. As illustrated in Fig. 2 , U-Net + + produced the closest visual agreement with the ground truth masks, not only preserving vessel continuity but also accurately capturing subtle variations in vessel thickness and fine anatomical details within narrower branches. This superior delineation aligns with its consistently higher per-image Dice scores (Table 1 ), reflecting robust overlap across both simple and complex test cases. U-Net also delivered reliable segmentations, effectively reconstructing the main coronary vessels. However, closer inspection revealed that its output occasionally displayed small discontinuities along vessel edges and, in some instances, slightly underestimated vessel diameters. These errors, although minor, demonstrate why its Dice scores trailed slightly behind those of U-Net++. In contrast, SegFormer, showed strong detection of major coronary structures, but its predictions often produced thicker boundaries compared to ground truth. In certain cases, this led to over-segmentation, particularly in areas with dense anatomical detail or overlapping vessels. While these outputs remained clinically interpretable, the reduced precision at vessel boundaries was reflected in SegFormer’s lower Dice averages. Taken together, these visual comparisons underscore that all three models performed strongly, but U-Net + + offered the most consistent balance between vessel completeness, anatomical fidelity, and boundary precision—qualities that are essential for downstream tasks such as stenosis assessment and treatment planning. Table 1 Per-image Dice scores on stenosis dataset. Image ID U-Net U-Net++ SegFormer 1 0.66 0.82 0.76 2 0.71 0.67 0.41 3 0.46 0.84 0.55 4 0.73 0.83 0.73 5 0.35 0.54 0.22 6 0.50 0.53 0.48 7 0.67 0.70 0.66 8 0.80 0.70 0.74 Average 0.61 0.70 0.57 4.1.2. Inference Time Evaluation Inference efficiency is critical for real-time deployment. Table 2 summarizes average prediction times per image. SegFormer was the fastest (~ 0.05 s/image), U-Net + + was moderately efficient (~ 0.18 s/image), and U-Net was slowest (~ 0.72 s/image). This suggests that while U-Net + + excels in accuracy, SegFormer offers clear advantages in speed for time-sensitive clinical environments. Table 2 Inference times on stenosis dataset (mean ± SD). Model Mean (s) Std (s) U-Net 0.72 0.15 U-Net++ 0.18 0.04 SegFormer 0.05 0.01 3.2. Results on SYNTAX Dataset 3.2.1. Segmentation Performance Analysis The SYNTAX dataset was more challenging due to greater anatomical variability and branching complexity. Validation accuracy decreased slightly: U-Net + + achieved the highest (98.13%), followed by U-Net (98.04%) and SegFormer (97.00%). Table 3 shows the per-image Dice scores for representative SYNTAX test images. Again, U-Net + + consistently outperformed the other models, particularly in preserving fine vessel structures. U-Net remained competitive, while SegFormer exhibited more variability, especially in complex branching regions. On SYNTAX subset, qualitative examples (Fig. 3 ) provide further insight into the performance differences across the three models. U-Net + + generated the most faithful reconstructions of vessel branching, consistently preserving continuity and capturing even the smaller peripheral branches that are often missed in automated segmentation. This ability to retain fine anatomical detail explains why U-Net + + achieved the highest per-image Dice scores (Table 3 ), confirming that its nested skip connections are particularly effective for complex coronary structures. U-Net also provided reliable segmentation, producing accurate delineations of the main coronary vessels. However, in several cases, its predictions displayed minor under-segmentation, particularly at the terminal ends of small branches, which slightly reduced overlap with the ground truth. On the other hand, SegFormer maintained strong detection of the primary vessels but tended to over-segment in regions of vessel overlap and bifurcations. This often resulted in thicker vessel boundaries and occasional inclusion of surrounding background pixels. While such predictions still yielded clinically useful vessel maps, they introduced variability that was reflected in SegFormer’s comparatively lower Dice averages. Taken together, these qualitative observations reinforce the quantitative evidence that U-Net + + strikes the best balance between vessel completeness and precision, U-Net serves as a dependable but slightly less detailed model, and SegFormer offers speed advantages at the cost of reduced boundary fidelity. Table 3 Per-image Dice scores on SYNTAX dataset. Image ID U-Net U-Net++ SegFormer 1 0.91 0.92 0.70 2 0.69 0.70 0.58 3 0.77 0.80 0.48 4 0.41 0.74 0.52 5 0.82 0.72 0.49 6 0.68 0.60 0.49 7 0.75 0.80 0.57 8 0.84 0.75 0.60 Average 0.73 0.75 0.55 3.2.2. Inference Time Evaluation Inference time results on the SYNTAX dataset differed from stenosis. As summarized in Table 4 , U-Net remained relatively consistent (~ 0.80 s/image), U-Net + + slowed considerably (~ 1.86 s/image), and SegFormer again proved most efficient (~ 0.18 s/image). This indicates that dataset complexity impacts U-Net + + disproportionately, while SegFormer maintains speed irrespective of vessel complexity. Table 4 Inference times on SYNTAX dataset (mean ± SD). Model Mean (s) Std (s) U-Net 0.80 0.2 U-Net++ 1.86 0.45 SegFormer 0.18 0.05 3.3. Comparative Discussion Across both datasets, U-Net + + consistently provided the best segmentation accuracy and Dice scores, highlighting its strength in capturing fine-grained vessel morphology and preserving boundary details. U-Net remained a robust baseline, delivering strong results with moderate computational efficiency. SegFormer achieved slightly lower accuracy but excelled in inference speed, showing remarkable consistency across datasets. The per-image Dice analysis underscores the importance of case-level evaluation. While average Dice scores were moderately high (ranging from 0.55 with SegFormer to 0.75 with U-Net++), variability across cases revealed differences in robustness. U-Net + + achieved the tightest distribution of Dice values, confirming its reliability across varied anatomies. SegFormer displayed greater variability, underscoring the trade-off between speed and precision. From a clinical perspective, U-Net + + would be most suitable in applications prioritizing segmentation accuracy (e.g., preoperative planning or research workflows), whereas SegFormer is more attractive for real-time settings such as catheterization labs, where rapid decision-making is critical. U-Net, although less advanced, offers a balanced trade-off and remains an interpretable, computationally manageable benchmark. 4. Conclusion This study presented a comparative analysis of three deep learning models—U-Net, U-Net++, and SegFormer—for coronary artery segmentation using the ARCADE dataset. Performance was assessed on both the stenosis and SYNTAX subsets in terms of segmentation accuracy, Dice similarity, and inference time. The results showed that U-Net + + consistently achieved the highest segmentation accuracy, particularly on the more complex SYNTAX subset, underscoring its effectiveness in capturing fine anatomical structures. SegFormer, although slightly less accurate, delivered the fastest inference times, making it highly attractive for real-time clinical applications. U-Net served as a reliable baseline, providing balanced performance between accuracy and computational efficiency. A key limitation of this study is the restricted availability of annotated coronary angiography datasets, which constrains broader generalizability. Future research should therefore focus on expanding annotated datasets in collaboration with healthcare providers and exploring strategies that reduce annotation dependence, such as semi-supervised or self-supervised learning. In brief, while U-Net++, U-Net, and SegFormer each demonstrated distinct strengths, no single model can be considered universally superior for coronary artery segmentation, as performance depends on dataset complexity, clinical context, and deployment requirements. Several promising deep learning approaches—including attention-based models, hybrid CNN–Transformer architectures, and lightweight frameworks—have not yet been systematically explored in this domain. Building on our comparative findings, future studies should evaluate a wider range of architectures, leverage larger and more diverse datasets, and adopt advanced training paradigms. Such efforts are expected to yield more robust, efficient, and clinically deployable solutions for coronary artery segmentation, ultimately enhancing decision support in cardiovascular imaging. Abbreviations 2D: Two-dimensional; 3D: Three-dimensional; ARCADE: Automatic Region-based Coronary Artery Disease Evaluation; CAD: Coronary artery disease; CVD: Cardiovascular disease; LMICs: Low- and middle-income countries; XCA: X-ray coronary angiography; CTA: Computed tomography angiography; CCTA: Coronary computed tomography angiography; QCA: Quantitative coronary angiography; SYNTAX: Synergy between PCI with Taxus and Cardiac Surgery score; PCI: Percutaneous coronary intervention; DL: Deep learning; CNN: Convolutional neural network; MLP: Multilayer perceptron; FFN: Feed-forward network; ReLU: Rectified linear unit; GPU: Graphics processing unit; DICOM: Digital Imaging and Communications in Medicine; JSON: JavaScript Object Notation; DSC: Dice similarity coefficient; F1: F1-score; TP: True positive; TN: True negative; FP: False positive; FN: False negative. Declarations Acknowledgements Not applicable Authors’ contributions BU wrote the code, performed the analyses, and drafted the manuscript. MD carried out the data preprocessing. YU and TI contributed to data management and organization. EN assisted in manuscript preparation. FU contributed to the medical imaging analysis. FM supported the technical aspects of the work. RNM and PSM provided additional input and critical feedback. JD designed the study and proofread the manuscript. All authors read and approved of the final manuscript. Funding Not applicable Availability of data and materials The datasets analyzed in the current study are publicly available from the referenced sources. The code used to support the findings of this study is available from the corresponding author upon reasonable request. Ethics Approval and Consent to Participate Not applicable Consent for Publication Not applicable. Competing Interests The authors declare that they have no competing interests. Author details College of Engineering, Carnegie Mellon University Africa, Kigali, Rwanda African Centre of Excellence in Data Science, University of Rwanda, Kigali, Rwanda College of Medicine and Health Sciences, University of Rwanda, Kigali, Rwanda College of Sciences and Technology, University of Rwanda, Kigali, Rwanda Ircad Africa, Kigali, Rwanda Department of Computer Science, Abilene Christian University, Texas, Abilene, United States Goizueta Business School, Emory University, Atlanta, GA 30322, United States References Pokharel B, Dhakal B. AJPHI I VOLUME 1 I 2024 C Coronary Artery Disease Demystified. Bansal A, Hiwale K. Updates in the Management of Coronary Artery Disease: A Review Article. Cureus. 2023 Dec 17; Al-Khlaiwi T, Habib SS, Bayoumy N, Al-Khliwi H, Meo SA. Identifying risk factors and mortality rate of premature coronary artery disease in young Saudi population. Sci Rep. 2024 Dec 1;14(1). Li Z, Lin L, Wu H, Yan L, Wang H, Yang H, et al. Global, Regional, and National Death, and Disability-Adjusted Life-Years (DALYs) for Cardiovascular Disease in 2017 and Trends and Risk Analysis From 1990 to 2017 Using the Global Burden of Disease Study and Implications for Prevention. Front Public Health. 2021 Oct 29;9. Unlu O, Fahed AC. Machine Learning in Invasive and Noninvasive Coronary Angiography. Vol. 25, Current Atherosclerosis Reports. Springer; 2023. p. 1025–33. Bansal A, Hiwale K. Updates in the Management of Coronary Artery Disease: A Review Article. Cureus. 2023 Dec 17; Yan J, Tian J, Yang H, Han G, Liu Y, He H, et al. A clinical decision support system for predicting coronary artery stenosis in patients with suspected coronary heart disease. Comput Biol Med. 2022 Dec 1;151. Rodriguez C, Pappas L, Le Hong Q, Baquero L, Nagel E. Cardiac imaging for the detection of ischemia: current status and future perspectives. Vol. 22, Expert Review of Medical Devices. Taylor and Francis Ltd.; 2025. p. 581–94. Danilov V V., Klyshnikov KY, Gerget OM, Kutikhin AG, Ganyukov VI, Frangi AF, et al. Real-time coronary artery stenosis detection based on modern neural networks. Sci Rep. 2021 Dec 1;11(1). Mézquita AJV, Biavati F, Falk V, Alkadhi H, Hajhosseiny R, Maurovich-Horvat P, et al. Clinical quantitative coronary artery stenosis and coronary atherosclerosis imaging: a Consensus Statement from the Quantitative Cardiovascular Imaging Study Group. Vol. 20, Nature Reviews Cardiology. Nature Research; 2023. p. 696–714. Gurav A, Revaiah PC, Tsai TY, Miyashita K, Tobe A, Oshima A, et al. Coronary angiography: a review of the state of the art and the evolution of angiography in cardio therapeutics. Vol. 11, Frontiers in Cardiovascular Medicine. Frontiers Media SA; 2024. World Health Organization (WHO). Cardiovascular diseases (CVDs) [Internet]. 2022 [cited 2025 Aug 23]. Available from: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds) Centers for Disease Control and Prevention (CDC). About Coronary Artery Disease (CAD). 2025 May. Nishi T, Yamashita R, Imura S, Tateishi K, Kitahara H, Kobayashi Y, et al. Deep learning-based intravascular ultrasound segmentation for the assessment of coronary artery disease. Int J Cardiol. 2021 Jun 15;333:55–9. Du T, Xie L, Zhang H, Liu X, Wang X, Chen D, et al. Training and validation of a deep learning architecture for the automatic analysis of coronary angiography. EuroIntervention. 2021 May 1;17(1):32–40. Zhu X, Cheng Z, Wang S, Chen X, Lu G. Coronary angiography image segmentation based on PSPNet. Comput Methods Programs Biomed. 2021 Mar 1;200. Pan LS, Li CW, Su SF, Tay SY, Tran QV, Chan WP. Coronary artery segmentation under class imbalance using a U-Net based architecture on computed tomography angiography images. Sci Rep. 2021 Dec 1;11(1). Cheung WK, Bell R, Nair A, Menezes LJ, Patel R, Wan S, et al. A Computationally Efficient Approach to Segmentation of the Aorta and Coronary Arteries Using Deep Learning. IEEE Access. 2021;9:108873–88. Tian F, Gao Y, Fang Z, Gu J. Automatic coronary artery segmentation algorithm based on deep learning and digital image processing. Applied Intelligence. 2021 Dec 1;51(12):8881–95. Balaji A, Kelsey LJ, Majeed K, Schultz CJ, Doyle BJ, Doyle B. Coronary Artery Segmentation from Intravascular Optical Coherence Tomography Using Deep Capsules. Jiang Z, Ou C, Qian Y, Rehan R, Yong A. Coronary vessel segmentation using multiresolution and multiscale deep learning. Inform Med Unlocked. 2021 Jan 1;24. Li Y, Wu Y, He J, Jiang W, Wang J, Peng Y, et al. Automatic coronary artery segmentation and diagnosis of stenosis by deep learning based on computed tomographic coronary angiography. Eur Radiol. 2022 Sep 1;32(9):6037–45. Jávorszky N, Homonnay B, Gerstenblith G, Bluemke D, Kiss P, Török M, et al. Deep learning–based atherosclerotic coronary plaque segmentation on coronary CT angiography. Eur Radiol. 2022 Oct 1;32(10):7217–26. Li Y, Wu Y, He J, Jiang W, Wang J, Peng Y, et al. Automatic coronary artery segmentation and diagnosis of stenosis by deep learning based on computed tomographic coronary angiography. Eur Radiol. 2022 Sep 1;32(9):6037–45. Dong C, Xu S, Li Z. A novel end-to-end deep learning solution for coronary artery segmentation from CCTA. Med Phys. 2022 Nov 1;49(11):6945–59. Dong C, Xu S, Li Z. A novel end-to-end deep learning solution for coronary artery segmentation from CCTA. Med Phys. 2022 Nov 1;49(11):6945–59. Algarni M, Al-Rezqi A, Saeed F, Alsaeedi A, Ghabban F. Multi-constraints based deep learning model for automated segmentation and diagnosis of coronary artery disease in X-ray angiographic images. PeerJ Comput Sci. 2022;8. Shen Y, Chen Z, Tong J, Jiang N, Ning Y. DBCU-Net: deep learning approach for segmentation of coronary angiography images. International Journal of Cardiovascular Imaging. 2023 Aug 1;39(8):1571–9. Serrano-Anton B, Otero-Cacho A, Lopez-Otero D, Diaz-Fernandez B, Bastos-Fernandez M, Perez-Munuzuri V, et al. Coronary Artery Segmentation Based on Transfer Learning and UNet Architecture on Computed Tomography Coronary Angiography Images. IEEE Access. 2023;11:75484–96. Kaba Ş, Haci H, Isin A, Ilhan A, Conkbayir C. The Application of Deep Learning for the Segmentation and Classification of Coronary Arteries. Diagnostics. 2023 Jul 1;13(13). Liu X, Wang X, Chen D, Zhang H. Automatic Quantitative Coronary Analysis Based on Deep Learning. Applied Sciences (Switzerland). 2023 Mar 1;13(5). Zhang X, Zhang B, Zhang F. Stenosis Detection and Quantification of Coronary Artery Using Machine Learning and Deep Learning. Vol. 75, Angiology. SAGE Publications Inc.; 2024. p. 405–16. Li S, Fan Y. Coronary Artery Segmentation in X-ray Angiography Based on Deep Learning Approach. Nannini G, Saitta S, Baggiano A, Maragna R, Mushtaq S, Pontone G, et al. A fully automated deep learning approach for coronary artery segmentation and comprehensive characterization. APL Bioeng. 2024 Mar 1;8(1). Jiang W, Li Y, Jia Y, Feng Y, Yi Z, Wang J, et al. Segmentation of coronary artery based on discriminative frequency learning and coronary-geometric refinement. Comput Biol Med. 2024 Oct 1;181. Muthusamy CD, Murugesh R. Integrated deep learning approach for automatic coronary artery segmentation and classification on computed tomographic coronary angiography. Network Modeling Analysis in Health Informatics and Bioinformatics. 2024 Dec 1;13(1). Wang L, Zhang X, Tian C, Chen S, Deng Y, Liao X, et al. PlaqueNet: deep learning enabled coronary artery plaque segmentation from coronary computed tomography angiography. Vis Comput Ind Biomed Art. 2024 Dec 1;7(1). Mahendiran T, Thanou D, Senouf O, Jamaa Y, Fournier S, De Bruyne B, et al. AngioPy Segmentation: An open-source, user-guided deep learning tool for coronary artery segmentation. Int J Cardiol. 2025 Jan 1;418. Morshedul Abedin AJ, Sarmun R, Mushtak A, Bin Mohamed Ali MS, Hasan A, Suganthan PN, et al. Enhanced coronary artery segmentation and stenosis detection: Leveraging novel deep learning techniques. Biomed Signal Process Control. 2025 Nov 1;109. Shrivastava P, Kashikar S, Parihar PH, Kasat P, Bhangale P, Shrivastava P. A systematic review on deep learning-enabled coronary CT angiography for plaque and stenosis quantification and cardiac risk prediction. Eur J Radiol Open. 2025 Jun 1;14. Alskaf E, Dutta U, Scannell CM, Chiribiri A. Deep learning applications in coronary anatomy imaging: a systematic review and meta-analysis. Available from: https://jmai.amegroups.com/article/ Gharleghi R, Chen N, Sowmya A, Beier S. Towards automated coronary artery segmentation: A systematic review. Vol. 225, Computer Methods and Programs in Biomedicine. Elsevier Ireland Ltd; 2022. Tu L, Deng Y, Chen Y, Luo Y. Accuracy of deep learning in the differential diagnosis of coronary artery stenosis: a systematic review and meta-analysis. BMC Med Imaging. 2024 Dec 1;24(1). Alnasser TN, Abdulaal L, Maiter A, Sharkey M, Dwivedi K, Salehi M, et al. Advancements in cardiac structures segmentation: a comprehensive systematic review of deep learning in CT imaging. Vol. 11, Frontiers in Cardiovascular Medicine. Frontiers Media SA; 2024. Wang L, Liang D, Yin X, Qiu J, Yang Z, Xing J, et al. Coronary artery segmentation in angiographic videos utilizing spatial-temporal information. BMC Med Imaging. 2020 Sep 24;20(1):110. Maxim Popov AANZAAASTA et al. ARCADE: Automatic Region-based Coronary Artery Disease diagnostics using x-ray angiography imagEs Dataset. 2023; Popov M, Amanturdieva A, Zhaksylyk N, Alkanov A, Saniyazbekov A, Aimyshev T, et al. Dataset for Automatic Region-based Coronary Artery Disease Diagnostics Using X-Ray Angiography Images. Sci Data. 2024;11(1):1–9. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2015;9351:234–41. Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation BT - Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Miccai. 2018;11045(2018):3–11. Xie E, Wang W, Yu Z, Anandkumar A, Alvarez JM, Luo P. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. Adv Neural Inf Process Syst. 2021;15(NeurIPS):12077–90. Kaba Ş, Haci H, Isin A, Ilhan A, Conkbayir C. The Application of Deep Learning for the Segmentation and Classification of Coronary Arteries. Diagnostics. 2023 Jul 1;13(13). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 15 Oct, 2025 Reviewers agreed at journal 29 Sep, 2025 Reviewers agreed at journal 24 Sep, 2025 Reviewers invited by journal 22 Sep, 2025 Editor invited by journal 28 Aug, 2025 Editor assigned by journal 28 Aug, 2025 Submission checks completed at journal 28 Aug, 2025 First submitted to journal 27 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7467606","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":523763868,"identity":"6b9dec71-ea44-4198-a449-50502e87d2dc","order_by":0,"name":"Benny Uhoranishema","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5UlEQVRIiWNgGAWjYBACAwYeMM3Dz8DcAKQloOJsRGiRbGAkUQuDwQGwFhjAo8VcIvfg44qaOzLGxw+2Sd3cYcHAP+2MAcOHssM4tVjOyEs2PHPsGY/ZmcQ26dwzEgwSt3MMGGecw63F4EaOmWQD22EeswMgLW0SDAbSOQbMvG2EtPw7zGPc/xBJy19CWhrbDvMYSCDbwohHi2XPG2PDxr7DPBI3HjZbA7XwSNxOKzjYcy4dpxZz9hzDhw3fDtvz9ycfvJ3bVifHPzt544MfZdY4tWAAcDQdIF79KBgFo2AUjAJsAAC/mFAuf0fpvAAAAABJRU5ErkJggg==","orcid":"","institution":"Carnegie Mellon University Africa","correspondingAuthor":true,"prefix":"","firstName":"Benny","middleName":"","lastName":"Uhoranishema","suffix":""},{"id":523763869,"identity":"56f7cf71-a368-48df-a0e5-fc4efca946ea","order_by":1,"name":"Mediatrice Dusenge","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Mediatrice","middleName":"","lastName":"Dusenge","suffix":""},{"id":523763870,"identity":"17922a19-2973-4297-a754-f2e97af4d930","order_by":2,"name":"Yollanda Umutoni","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Yollanda","middleName":"","lastName":"Umutoni","suffix":""},{"id":523763871,"identity":"91a9337a-c7be-44b9-9de0-191a091ee78d","order_by":3,"name":"Theophilla Igihozo","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Theophilla","middleName":"","lastName":"Igihozo","suffix":""},{"id":523763872,"identity":"2ef59bb5-90ef-4a9f-865e-920c5d88c963","order_by":4,"name":"Emmanuella Nuwayo","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Emmanuella","middleName":"","lastName":"Nuwayo","suffix":""},{"id":523763873,"identity":"f7ef1679-5d04-4172-a560-333d1acbc9ee","order_by":5,"name":"Florien Ujemurwego","email":"","orcid":"","institution":"Ircad Africa","correspondingAuthor":false,"prefix":"","firstName":"Florien","middleName":"","lastName":"Ujemurwego","suffix":""},{"id":523763874,"identity":"29c36f90-85a9-4a1a-abd2-873bfae4084e","order_by":6,"name":"Fabrice Mpozenzi","email":"","orcid":"","institution":"Abilene Christian University","correspondingAuthor":false,"prefix":"","firstName":"Fabrice","middleName":"","lastName":"Mpozenzi","suffix":""},{"id":523763875,"identity":"aaa56e0a-3d56-4d1f-80e4-97c3e4f8bac9","order_by":7,"name":"Robert Ngabo Mugisha","email":"","orcid":"","institution":"Emory University","correspondingAuthor":false,"prefix":"","firstName":"Robert","middleName":"Ngabo","lastName":"Mugisha","suffix":""},{"id":523763876,"identity":"023a4035-0d78-46f0-9502-d87d9557ee34","order_by":8,"name":"Prince Shema Musonerwa","email":"","orcid":"","institution":"Emory University","correspondingAuthor":false,"prefix":"","firstName":"Prince","middleName":"Shema","lastName":"Musonerwa","suffix":""},{"id":523763877,"identity":"36f6d3ae-bf4b-4a1c-9de5-7d9148daa3e7","order_by":9,"name":"Jean De Dieu Niyonteze","email":"","orcid":"","institution":"Emory University","correspondingAuthor":false,"prefix":"","firstName":"Jean","middleName":"De Dieu","lastName":"Niyonteze","suffix":""}],"badges":[],"createdAt":"2025-08-27 04:09:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7467606/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7467606/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92735908,"identity":"4ab11ee4-91c6-49eb-9c78-4e93ef0427f9","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":591306,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/79d8e3816dd1ba6f8090e4a9.docx"},{"id":92735907,"identity":"5c20c311-bfa3-4895-8301-ed7521bdacd3","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11451,"visible":true,"origin":"","legend":"","description":"","filename":"cc7bec2b1a5844de968ba77376d5081a.json","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/94b816bd48ec8be1dad386c9.json"},{"id":92735916,"identity":"e3304cdd-8e25-4ca5-944e-eb7c4c3ada11","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115908,"visible":true,"origin":"","legend":"","description":"","filename":"cc7bec2b1a5844de968ba77376d5081a1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/9e4ad439d6fb99d3660c3041.xml"},{"id":92738075,"identity":"5a301d93-968a-4287-ac31-932c01f645de","added_by":"auto","created_at":"2025-10-03 16:48:35","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70616,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/cfaa5241201984f0015a2e22.png"},{"id":92737104,"identity":"6e4ccef7-fca5-4fa4-9c42-615cc01c4438","added_by":"auto","created_at":"2025-10-03 16:40:35","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119284,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/c81f75c0cca041b7c0c132d1.png"},{"id":92735914,"identity":"b4539088-3f43-4259-85a5-67c67a3e773f","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":189304,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/acb720b513f9a2b2079373ae.png"},{"id":92735917,"identity":"72b68f2a-b86f-40ca-96f0-3020be7c56ab","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116611,"visible":true,"origin":"","legend":"","description":"","filename":"cc7bec2b1a5844de968ba77376d5081a1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/ab44bdc70216c361712fd0c8.xml"},{"id":92735915,"identity":"f0f6e60f-2f3d-44be-ad2e-95a2610514b1","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"html","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125956,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/b0242ea0704d98b8802d301a.html"},{"id":92735912,"identity":"7fa10d12-3809-46f2-aaa4-c5c69e366787","added_by":"auto","created_at":"2025-10-03 16:32:35","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":134661,"visible":true,"origin":"","legend":"\u003cp\u003eTraining performance of U-Net++ on the stenosis dataset: (a) training and validation loss, and (b) training and validation accuracy.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/2bb2b68fa410f3fe0e872e16.png"},{"id":92737101,"identity":"b83bc5c8-0433-4e0a-bc33-e248ccf44ce0","added_by":"auto","created_at":"2025-10-03 16:40:35","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":74211,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative segmentation results on the stenosis dataset: comparison of ground truth and predictions from U-Net, U-Net++, and SegFormer.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/111201af10492eda445809e6.jpeg"},{"id":92738074,"identity":"edeef157-636b-4289-9a46-ae294cf9642a","added_by":"auto","created_at":"2025-10-03 16:48:35","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":113957,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative segmentation results on the SYNTAX dataset: comparison of ground truth and predictions from U-Net, U-Net++, and SegFormer.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/aedc2a8e5d2ac32525bebd07.jpeg"},{"id":92739007,"identity":"a583410c-609a-444c-9e94-b15356c1ca02","added_by":"auto","created_at":"2025-10-03 16:56:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1034893,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7467606/v1/9e024a2b-f58f-4271-b740-a33889534071.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparative Analysis of Deep Learning Models for Coronary Artery Segmentation: Performance and Inference Time Evaluation","fulltext":[{"header":"1. Background","content":"\u003cp\u003eCoronary artery disease (CAD) is one of the leading causes of cardiovascular mortality worldwide, contributing substantially to the global burden of cardiovascular diseases (\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). Accurate detection and localization of coronary artery stenosis are therefore crucial for guiding effective clinical decision-making. Coronary angiography continues to serve as the gold standard for diagnosing CAD; however, manual interpretation of angiograms is time-consuming, subject to observer variability, and prone to diagnostic errors (\u003cspan additionalcitationids=\"CR6 CR7 CR8 CR9 CR10\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe scale of this challenge is underscored by global and regional statistics. Cardiovascular diseases (CVDs) remain the leading cause of death worldwide, contributing to a substantial number of fatalities and disabilities. In 2021, the World Health Organization (WHO) reported that CVDs were responsible for 20.5\u0026nbsp;million deaths, representing nearly one-third of all global deaths. Although once regarded as conditions of affluence, CVDs now disproportionately affect low- and middle-income countries (LMICs), where more than three-quarters of related deaths occur (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). On the other hand, CAD is the most common type of heart disease caused by plaque buildup in the walls of the arteries that supply blood to the heart (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Therefore, this progressive accumulation of plaque leads to coronary artery stenosis, a narrowing of the arterial lumen that restricts blood flow and can trigger angina, myocardial infarction, and other serious cardiovascular events.\u003c/p\u003e\u003cp\u003eDespite these challenges in cardiovascular diseases particularly CAD, Deep Learning (DL) has emerged as a highly promising approach for coronary artery segmentation. It offers precise, reproducible, and automated solutions for both coronary artery segmentation and stenosis localization, thereby holding great potential to transform traditional clinical practice.\u003c/p\u003e\u003cp\u003eIn 2021, researchers in Ref. (\u003cspan additionalcitationids=\"CR15 CR16 CR17 CR18 CR19 CR20\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e) made significant progress in applying deep learning to coronary artery segmentation and analysis. Existing architectures such as U-Net and Dense U-Net were widely adopted, while several new models were introduced, including multiresolution and multiscale U-Net, cascaded V-shaped networks, DeepCap (capsule networks), and DeepDiscern (dual-network system). These approaches achieved notable improvements in segmentation accuracy, robustness against class imbalance, and computational efficiency, with Dice scores reaching up to 0.97, thereby demonstrating strong potential for coronary artery assessment.\u003c/p\u003e\u003cp\u003eNext, from 2022 to 2024, studies consistently advanced coronary artery segmentation by refining established models such as U-Net, 3D U-Net, and DenseNet, while also introducing a few new architectures including Di-VNet, DBCU-Net, PlaqueNet, and GhostNet-U-Net (\u003cspan additionalcitationids=\"CR23 CR24 CR25 CR26 CR27 CR28 CR29 CR30 CR31 CR32 CR33 CR34 CR35 CR36\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e). These efforts steadily improved segmentation accuracy, efficiency, and clinical applicability across modalities such as CCTA (Coronary Computed Tomography Angiography), XCA, and quantitative angiography.\u003c/p\u003e\u003cp\u003eMore importantly, while drafting this paper in mid-2025, Mahendiran et al. published AngioPy Segmentation: An open-source, user-guided deep learning tool for coronary artery segmentation, which integrates user-defined ground-truth points to improve segmentation accuracy and reduce the need for manual correction. The model achieved an average F1 score of 0.927 on internal datasets and 0.924 on external validation, demonstrating excellent agreement with established quantitative coronary angiography (QCA) systems and highlighting its potential to enhance clinical accuracy and efficiency (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e). Besides, Abedin et al. introduced \u003cem\u003eDenseSelfU-Net\u003c/em\u003e and \u003cem\u003eDenseSelfMA-Net\u003c/em\u003e, two Self-Organizing Neural Network (Self-ONN) enhanced frameworks designed for coronary artery segmentation and stenosis localization. On the ARCADE challenge dataset, DenseSelfU-Net achieved a Dice score of 90.35% for full artery segmentation, while DenseSelfMA-Net configurations reached up to 60.92% Dice for stenosis localization, outperforming prior methods such as StenU-Net and SSASS (\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eStarting in 2022, researchers began publishing comprehensive and systematic reviews on deep learning applications for coronary artery imaging. These reviews, including meta-analyses, have primarily examined CCTA and angiography, systematically assessing deep learning applications for coronary artery segmentation, stenosis detection, plaque quantification, and risk prediction. Collectively, they highlight the high diagnostic accuracy of DL models, their growing clinical potential, and the need for multicenter validation and standardized methodologies (\u003cspan additionalcitationids=\"CR41 CR42 CR43\" citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eOne study by Wang et al. proposed a 3D\u0026ndash;2D convolutional network for coronary artery segmentation in angiographic videos, which not only improved vessel detail preservation compared to U-Net and CE-Net, but also achieved inference times of 0.054\u0026ndash;0.061 seconds per frame, thereby meeting the requirements for real-time clinical use (\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWhile numerous studies have focused primarily on improving segmentation accuracy, only a very limited number have explicitly evaluated inference time. To the best of our knowledge, this is the first study to provide a comparative analysis of deep learning models for coronary artery segmentation by jointly assessing both performance and inference efficiency, an aspect that is essential for real-time clinical deployment.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eThe proposed system employs a DL\u0026ndash;based pipeline for coronary artery segmentation from computed tomography angiography (CTA) images. The workflow comprises three main stages: data preprocessing, model training, and performance evaluation. Preprocessing includes image enhancement techniques to improve vessel visibility and facilitate more accurate feature learning. Deep learning models\u0026mdash;specifically U-Net, U-Net++, and SegFormer\u0026mdash;were trained to perform pixel-level segmentation of coronary arteries, enabling precise extraction of vessel structures critical for cardiovascular disease assessment.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Dataset\u003c/h2\u003e\u003cp\u003eThis study utilized the publicly available \u003cem\u003eAutomatic Coronary Artery Disease Evaluation (ARCADE)\u003c/em\u003e dataset (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e). The dataset is organized into two main objectives: (i) vessel branch classification according to the SYNTAX methodology (syntax folder), and (ii) stenosis detection (stenosis folder). Each objective contains three subsets (train, val, and test), with corresponding images in \u0026ldquo;.png\u0026rdquo; format extracted from DICOM recordings, and annotation files provided in \u0026ldquo;.JSON\u0026rdquo; format. The annotation files include image metadata (ID, width, height, file name), categorical labels (26 SYNTAX-based vessel regions or stenosis classes), and detailed mask information including segmentation points, bounding box coordinates, and area. The dataset was carefully annotated by medical experts to support research on automated risk assessment and diagnosis of coronary artery disease (\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Segmentation Models\u003c/h2\u003e\u003cdiv id=\"Sec5\" class=\"Section3\"\u003e\u003ch2\u003e3.2.1. U-Net\u003c/h2\u003e\u003cp\u003eThe U-Net architecture follows a symmetric encoder\u0026ndash;decoder structure with skip connections, enabling precise localization in biomedical image segmentation tasks. The contracting path consists of repeated applications of two 3\u0026times;3 convolutions (unpadded), each followed by a rectified linear unit (ReLU), and a 2\u0026times;2 max-pooling operation with stride 2 for down sampling, doubling the number of feature channels at each step. The expansive path mirrors this process, where each step involves sampling of the feature map, followed by a 2\u0026times;2 up-convolution that halves the number of feature channels, concatenation with the corresponding cropped feature map from the contracting path, and two successive 3\u0026times;3 convolutions with ReLU activation. Cropping is applied to compensate for border pixel loss during convolution. At the final layer, a 1\u0026times;1 convolution maps the 64-component feature vectors to the desired number of classes. In total, the network comprises 23 convolutional layers and exhibits the characteristic U-shaped topology (\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\u003ch2\u003e3.2.2. UNet++\u003c/h2\u003e\u003cp\u003eU-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;was proposed by Zhou et al. as an extension of the original U-Net for medical image segmentation. It modifies the baseline U-Net by introducing a dense network of skip connections between the encoder and decoder paths. The architecture links the encoder and decoder through a series of nested, dense skip pathways, which progressively refine feature maps before fusion. Unlike U-Net, where encoder features are directly transmitted to the decoder, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;maps encoder outputs to the decoder via dense convolutional blocks within redesigned skip pathways. This design reduces the semantic gap between encoder and decoder feature maps, ensuring that the feature representations at corresponding levels are more consistent. As a result, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;achieves better alignment between low-level spatial details and high-level semantic information, thereby improving segmentation performance (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\u003ch2\u003e3.2.3. SegFormer\u003c/h2\u003e\u003cp\u003eSegFormer is a simple yet efficient semantic segmentation framework that combines a hierarchical Transformer encoder with a lightweight multilayer perceptron (MLP) decoder. The encoder produces multistage features without relying on positional encoding, which improves robustness when the testing resolution differs from training. To further enhance efficiency, SegFormer reshapes and projects the key matrix in self-attention, significantly reducing computational cost. The decoder employs a Mix-FFN block, which integrates depth-wise 3\u0026times;3 convolution into the MLP to efficiently inject positional information while keeping the parameter count low. By avoiding complex decoders, SegFormer achieves an excellent balance between accuracy and efficiency in semantic segmentation (\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Evaluation Metrics\u003c/h2\u003e\u003cp\u003eFor model evaluation, we employed standard segmentation metrics (\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAccuracy measures the fraction of correctly classified pixels and is defined as:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\frac{TP+TN}{TP+TN+FP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cem\u003eTP\u003c/em\u003e represents true positives, \u003cem\u003eTN\u003c/em\u003e true negatives, \u003cem\u003eFP\u003c/em\u003e false positives, and \u003cem\u003eFN\u003c/em\u003e false negatives. Accuracy is straightforward to interpret and suitable for balanced datasets; however, it may be misleading for imbalanced data, as it does not account for the relative impact of false positives and false negatives.\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:Dice\\:Score=\\frac{2TP}{2TP+FP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn terms of TP, FP, FN: Dice score referred to as the F1-score in machine learning, ranges from 0 to 1 and is particularly effective for image segmentation and object detection tasks, as it emphasizes the correct identification of positive cases. More importantly, it is widely used in medical image segmentation because it emphasizes the correct identification of positive (foreground) pixels and provides a more reliable measure of segmentation quality than accuracy alone.\u003c/p\u003e\u003cp\u003eOn the other hand, Dice measures the overlap between the predicted segmentation and the ground truth, where X represents the set of predicted positive pixels and Y represents the set of ground truth positive pixels.\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:Dice\\:Score=\\frac{2\\:.\\:\\:X\\cap\\:Y\\:}{\\left|X\\right|+\\left|Y\\right|}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cp\u003eThis section reports the segmentation performance of U-Net, U-Net++, and SegFormer on the stenosis and SYNTAX datasets. Performance was assessed using training and validation accuracy, per-image Dice similarity coefficients (DSC), and inference time. The Dice score, widely used in medical image segmentation, quantifies the overlap between predicted masks and ground truth annotations for each image. A value of 1 indicates perfect overlap, whereas 0 indicates no overlap. Reporting per-image Dice scores allows us to capture variability across cases and better understand model generalization. Moreover, all model training, validation, prediction, and inference time assessments were conducted on an NVIDIA GeForce RTX 4070 GPU.\u003c/p\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Results on Stenosis Dataset\u003c/h2\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.1.1. Segmentation Performance Analysis\u003c/h2\u003e\u003cp\u003eAll three models achieved very high segmentation accuracy on the stenosis dataset. U-Net and U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;both reached validation accuracies of 99.82%, while SegFormer slightly lagged at 99.15%. Representative training dynamics of U-Net++ (the best performer on both subsets) are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, demonstrating stable convergence with minimal overfitting.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the per-image Dice scores on a representative set of stenosis test images, along with the average across the dataset. U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;consistently produced the highest Dice values, reflecting superior overlap with the ground truth. U-Net followed closely, while SegFormer showed slightly lower scores, particularly in cases with complex vessel boundaries.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eBeyond the numerical results, qualitative analysis further confirmed the quantitative findings and offered a more intuitive understanding of model behavior. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;produced the closest visual agreement with the ground truth masks, not only preserving vessel continuity but also accurately capturing subtle variations in vessel thickness and fine anatomical details within narrower branches. This superior delineation aligns with its consistently higher per-image Dice scores (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), reflecting robust overlap across both simple and complex test cases. U-Net also delivered reliable segmentations, effectively reconstructing the main coronary vessels. However, closer inspection revealed that its output occasionally displayed small discontinuities along vessel edges and, in some instances, slightly underestimated vessel diameters. These errors, although minor, demonstrate why its Dice scores trailed slightly behind those of U-Net++. In contrast, SegFormer, showed strong detection of major coronary structures, but its predictions often produced thicker boundaries compared to ground truth. In certain cases, this led to over-segmentation, particularly in areas with dense anatomical detail or overlapping vessels. While these outputs remained clinically interpretable, the reduced precision at vessel boundaries was reflected in SegFormer\u0026rsquo;s lower Dice averages. Taken together, these visual comparisons underscore that all three models performed strongly, but U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;offered the most consistent balance between vessel completeness, anatomical fidelity, and boundary precision\u0026mdash;qualities that are essential for downstream tasks such as stenosis assessment and treatment planning.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePer-image Dice scores on stenosis dataset.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eImage ID\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eU-Net++\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSegFormer\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.66\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.67\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.41\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.46\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.55\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.35\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.22\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.48\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.67\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.66\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAverage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.61\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.70\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.57\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e4.1.2. Inference Time Evaluation\u003c/h2\u003e\u003cp\u003eInference efficiency is critical for real-time deployment. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes average prediction times per image. SegFormer was the fastest (~\u0026thinsp;0.05 s/image), U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;was moderately efficient (~\u0026thinsp;0.18 s/image), and U-Net was slowest (~\u0026thinsp;0.72 s/image). This suggests that while U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;excels in accuracy, SegFormer offers clear advantages in speed for time-sensitive clinical environments.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eInference times on stenosis dataset (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMean (s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eStd (s)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.15\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eU-Net++\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.04\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSegFormer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.01\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Results on SYNTAX Dataset\u003c/h2\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e3.2.1. Segmentation Performance Analysis\u003c/h2\u003e\u003cp\u003eThe SYNTAX dataset was more challenging due to greater anatomical variability and branching complexity. Validation accuracy decreased slightly: U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;achieved the highest (98.13%), followed by U-Net (98.04%) and SegFormer (97.00%).\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the per-image Dice scores for representative SYNTAX test images. Again, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;consistently outperformed the other models, particularly in preserving fine vessel structures. U-Net remained competitive, while SegFormer exhibited more variability, especially in complex branching regions.\u003c/p\u003e\u003cp\u003eOn SYNTAX subset, qualitative examples (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) provide further insight into the performance differences across the three models. U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;generated the most faithful reconstructions of vessel branching, consistently preserving continuity and capturing even the smaller peripheral branches that are often missed in automated segmentation. This ability to retain fine anatomical detail explains why U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;achieved the highest per-image Dice scores (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), confirming that its nested skip connections are particularly effective for complex coronary structures. U-Net also provided reliable segmentation, producing accurate delineations of the main coronary vessels. However, in several cases, its predictions displayed minor under-segmentation, particularly at the terminal ends of small branches, which slightly reduced overlap with the ground truth. On the other hand, SegFormer maintained strong detection of the primary vessels but tended to over-segment in regions of vessel overlap and bifurcations. This often resulted in thicker vessel boundaries and occasional inclusion of surrounding background pixels. While such predictions still yielded clinically useful vessel maps, they introduced variability that was reflected in SegFormer\u0026rsquo;s comparatively lower Dice averages. Taken together, these qualitative observations reinforce the quantitative evidence that U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;strikes the best balance between vessel completeness and precision, U-Net serves as a dependable but slightly less detailed model, and SegFormer offers speed advantages at the cost of reduced boundary fidelity.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePer-image Dice scores on SYNTAX dataset.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eImage ID\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eU-Net++\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSegFormer\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.70\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.58\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.48\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.52\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.68\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.57\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAverage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.73\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.75\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.55\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003ch2\u003e3.2.2. Inference Time Evaluation\u003c/h2\u003e\u003cp\u003eInference time results on the SYNTAX dataset differed from stenosis. As summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, U-Net remained relatively consistent (~\u0026thinsp;0.80 s/image), U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;slowed considerably (~\u0026thinsp;1.86 s/image), and SegFormer again proved most efficient (~\u0026thinsp;0.18 s/image). This indicates that dataset complexity impacts U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;disproportionately, while SegFormer maintains speed irrespective of vessel complexity.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eInference times on SYNTAX dataset (mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMean (s)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eStd (s)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eU-Net++\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.45\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSegFormer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Comparative Discussion\u003c/h2\u003e\u003cp\u003eAcross both datasets, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;consistently provided the best segmentation accuracy and Dice scores, highlighting its strength in capturing fine-grained vessel morphology and preserving boundary details. U-Net remained a robust baseline, delivering strong results with moderate computational efficiency. SegFormer achieved slightly lower accuracy but excelled in inference speed, showing remarkable consistency across datasets.\u003c/p\u003e\u003cp\u003eThe per-image Dice analysis underscores the importance of case-level evaluation. While average Dice scores were moderately high (ranging from 0.55 with SegFormer to 0.75 with U-Net++), variability across cases revealed differences in robustness. U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;achieved the tightest distribution of Dice values, confirming its reliability across varied anatomies. SegFormer displayed greater variability, underscoring the trade-off between speed and precision.\u003c/p\u003e\u003cp\u003eFrom a clinical perspective, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;would be most suitable in applications prioritizing segmentation accuracy (e.g., preoperative planning or research workflows), whereas SegFormer is more attractive for real-time settings such as catheterization labs, where rapid decision-making is critical. U-Net, although less advanced, offers a balanced trade-off and remains an interpretable, computationally manageable benchmark.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003eThis study presented a comparative analysis of three deep learning models\u0026mdash;U-Net, U-Net++, and SegFormer\u0026mdash;for coronary artery segmentation using the ARCADE dataset. Performance was assessed on both the stenosis and SYNTAX subsets in terms of segmentation accuracy, Dice similarity, and inference time.\u003c/p\u003e\u003cp\u003eThe results showed that U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;consistently achieved the highest segmentation accuracy, particularly on the more complex SYNTAX subset, underscoring its effectiveness in capturing fine anatomical structures. SegFormer, although slightly less accurate, delivered the fastest inference times, making it highly attractive for real-time clinical applications. U-Net served as a reliable baseline, providing balanced performance between accuracy and computational efficiency.\u003c/p\u003e\u003cp\u003eA key limitation of this study is the restricted availability of annotated coronary angiography datasets, which constrains broader generalizability. Future research should therefore focus on expanding annotated datasets in collaboration with healthcare providers and exploring strategies that reduce annotation dependence, such as semi-supervised or self-supervised learning.\u003c/p\u003e\u003cp\u003eIn brief, while U-Net++, U-Net, and SegFormer each demonstrated distinct strengths, no single model can be considered universally superior for coronary artery segmentation, as performance depends on dataset complexity, clinical context, and deployment requirements. Several promising deep learning approaches\u0026mdash;including attention-based models, hybrid CNN\u0026ndash;Transformer architectures, and lightweight frameworks\u0026mdash;have not yet been systematically explored in this domain. Building on our comparative findings, future studies should evaluate a wider range of architectures, leverage larger and more diverse datasets, and adopt advanced training paradigms. Such efforts are expected to yield more robust, efficient, and clinically deployable solutions for coronary artery segmentation, ultimately enhancing decision support in cardiovascular imaging.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e2D: Two-dimensional; 3D: Three-dimensional; ARCADE: Automatic Region-based Coronary Artery Disease Evaluation; CAD: Coronary artery disease; CVD: Cardiovascular disease; LMICs: Low- and middle-income countries; XCA: X-ray coronary angiography; CTA: Computed tomography angiography; CCTA: Coronary computed tomography angiography; QCA: Quantitative coronary angiography; SYNTAX: Synergy between PCI with Taxus and Cardiac Surgery score; PCI: Percutaneous coronary intervention; DL: Deep learning; CNN: Convolutional neural network; MLP: Multilayer perceptron; FFN: Feed-forward network; ReLU: Rectified linear unit; GPU: Graphics processing unit; DICOM: Digital Imaging and Communications in Medicine; JSON: JavaScript Object Notation; DSC: Dice similarity coefficient; F1: F1-score; TP: True positive; TN: True negative; FP: False positive; FN: False negative.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBU wrote the code, performed the analyses, and drafted the manuscript. MD carried out the data preprocessing. YU and TI contributed to data management and organization. EN assisted in manuscript preparation. FU contributed to the medical imaging analysis. FM supported the technical aspects of the work. RNM and PSM provided additional input and critical feedback. JD designed the study and proofread the manuscript. All authors read and approved of the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets analyzed in the current study are publicly available from the referenced sources. The code used to support the findings of this study is available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor details\u003c/strong\u003e\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eCollege of Engineering, Carnegie Mellon University Africa, Kigali, Rwanda\u003c/li\u003e\n \u003cli\u003eAfrican Centre of Excellence in Data Science, University of Rwanda, Kigali, Rwanda\u003c/li\u003e\n \u003cli\u003eCollege of Medicine and Health Sciences, University of Rwanda, Kigali, Rwanda\u003c/li\u003e\n \u003cli\u003eCollege of Sciences and Technology, University of Rwanda, Kigali, Rwanda\u003c/li\u003e\n \u003cli\u003eIrcad Africa, Kigali, Rwanda\u003c/li\u003e\n \u003cli\u003eDepartment of Computer Science, Abilene Christian University, Texas, Abilene, United States\u003c/li\u003e\n \u003cli\u003eGoizueta Business School, Emory University, Atlanta, GA 30322, United States \u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePokharel B, Dhakal B. AJPHI I VOLUME 1 I 2024 C Coronary Artery Disease Demystified. \u003c/li\u003e\n\u003cli\u003eBansal A, Hiwale K. Updates in the Management of Coronary Artery Disease: A Review Article. Cureus. 2023 Dec 17; \u003c/li\u003e\n\u003cli\u003eAl-Khlaiwi T, Habib SS, Bayoumy N, Al-Khliwi H, Meo SA. Identifying risk factors and mortality rate of premature coronary artery disease in young Saudi population. Sci Rep. 2024 Dec 1;14(1). \u003c/li\u003e\n\u003cli\u003eLi Z, Lin L, Wu H, Yan L, Wang H, Yang H, et al. Global, Regional, and National Death, and Disability-Adjusted Life-Years (DALYs) for Cardiovascular Disease in 2017 and Trends and Risk Analysis From 1990 to 2017 Using the Global Burden of Disease Study and Implications for Prevention. Front Public Health. 2021 Oct 29;9. \u003c/li\u003e\n\u003cli\u003eUnlu O, Fahed AC. Machine Learning in Invasive and Noninvasive Coronary Angiography. Vol. 25, Current Atherosclerosis Reports. Springer; 2023. p. 1025\u0026ndash;33. \u003c/li\u003e\n\u003cli\u003eBansal A, Hiwale K. Updates in the Management of Coronary Artery Disease: A Review Article. Cureus. 2023 Dec 17; \u003c/li\u003e\n\u003cli\u003eYan J, Tian J, Yang H, Han G, Liu Y, He H, et al. A clinical decision support system for predicting coronary artery stenosis in patients with suspected coronary heart disease. Comput Biol Med. 2022 Dec 1;151. \u003c/li\u003e\n\u003cli\u003eRodriguez C, Pappas L, Le Hong Q, Baquero L, Nagel E. Cardiac imaging for the detection of ischemia: current status and future perspectives. Vol. 22, Expert Review of Medical Devices. Taylor and Francis Ltd.; 2025. p. 581\u0026ndash;94. \u003c/li\u003e\n\u003cli\u003eDanilov V V., Klyshnikov KY, Gerget OM, Kutikhin AG, Ganyukov VI, Frangi AF, et al. Real-time coronary artery stenosis detection based on modern neural networks. Sci Rep. 2021 Dec 1;11(1). \u003c/li\u003e\n\u003cli\u003eM\u0026eacute;zquita AJV, Biavati F, Falk V, Alkadhi H, Hajhosseiny R, Maurovich-Horvat P, et al. Clinical quantitative coronary artery stenosis and coronary atherosclerosis imaging: a Consensus Statement from the Quantitative Cardiovascular Imaging Study Group. Vol. 20, Nature Reviews Cardiology. Nature Research; 2023. p. 696\u0026ndash;714. \u003c/li\u003e\n\u003cli\u003eGurav A, Revaiah PC, Tsai TY, Miyashita K, Tobe A, Oshima A, et al. Coronary angiography: a review of the state of the art and the evolution of angiography in cardio therapeutics. Vol. 11, Frontiers in Cardiovascular Medicine. Frontiers Media SA; 2024. \u003c/li\u003e\n\u003cli\u003eWorld Health Organization (WHO). Cardiovascular diseases (CVDs) [Internet]. 2022 [cited 2025 Aug 23]. Available from: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)\u003c/li\u003e\n\u003cli\u003eCenters for Disease Control and Prevention (CDC). About Coronary Artery Disease (CAD). 2025 May. \u003c/li\u003e\n\u003cli\u003eNishi T, Yamashita R, Imura S, Tateishi K, Kitahara H, Kobayashi Y, et al. Deep learning-based intravascular ultrasound segmentation for the assessment of coronary artery disease. Int J Cardiol. 2021 Jun 15;333:55\u0026ndash;9. \u003c/li\u003e\n\u003cli\u003eDu T, Xie L, Zhang H, Liu X, Wang X, Chen D, et al. Training and validation of a deep learning architecture for the automatic analysis of coronary angiography. EuroIntervention. 2021 May 1;17(1):32\u0026ndash;40. \u003c/li\u003e\n\u003cli\u003eZhu X, Cheng Z, Wang S, Chen X, Lu G. Coronary angiography image segmentation based on PSPNet. Comput Methods Programs Biomed. 2021 Mar 1;200. \u003c/li\u003e\n\u003cli\u003ePan LS, Li CW, Su SF, Tay SY, Tran QV, Chan WP. Coronary artery segmentation under class imbalance using a U-Net based architecture on computed tomography angiography images. Sci Rep. 2021 Dec 1;11(1). \u003c/li\u003e\n\u003cli\u003eCheung WK, Bell R, Nair A, Menezes LJ, Patel R, Wan S, et al. A Computationally Efficient Approach to Segmentation of the Aorta and Coronary Arteries Using Deep Learning. IEEE Access. 2021;9:108873\u0026ndash;88. \u003c/li\u003e\n\u003cli\u003eTian F, Gao Y, Fang Z, Gu J. Automatic coronary artery segmentation algorithm based on deep learning and digital image processing. Applied Intelligence. 2021 Dec 1;51(12):8881\u0026ndash;95. \u003c/li\u003e\n\u003cli\u003eBalaji A, Kelsey LJ, Majeed K, Schultz CJ, Doyle BJ, Doyle B. Coronary Artery Segmentation from Intravascular Optical Coherence Tomography Using Deep Capsules. \u003c/li\u003e\n\u003cli\u003eJiang Z, Ou C, Qian Y, Rehan R, Yong A. Coronary vessel segmentation using multiresolution and multiscale deep learning. Inform Med Unlocked. 2021 Jan 1;24. \u003c/li\u003e\n\u003cli\u003eLi Y, Wu Y, He J, Jiang W, Wang J, Peng Y, et al. Automatic coronary artery segmentation and diagnosis of stenosis by deep learning based on computed tomographic coronary angiography. Eur Radiol. 2022 Sep 1;32(9):6037\u0026ndash;45. \u003c/li\u003e\n\u003cli\u003eJ\u0026aacute;vorszky N, Homonnay B, Gerstenblith G, Bluemke D, Kiss P, T\u0026ouml;r\u0026ouml;k M, et al. Deep learning\u0026ndash;based atherosclerotic coronary plaque segmentation on coronary CT angiography. Eur Radiol. 2022 Oct 1;32(10):7217\u0026ndash;26. \u003c/li\u003e\n\u003cli\u003eLi Y, Wu Y, He J, Jiang W, Wang J, Peng Y, et al. Automatic coronary artery segmentation and diagnosis of stenosis by deep learning based on computed tomographic coronary angiography. Eur Radiol. 2022 Sep 1;32(9):6037\u0026ndash;45. \u003c/li\u003e\n\u003cli\u003eDong C, Xu S, Li Z. A novel end-to-end deep learning solution for coronary artery segmentation from CCTA. Med Phys. 2022 Nov 1;49(11):6945\u0026ndash;59. \u003c/li\u003e\n\u003cli\u003eDong C, Xu S, Li Z. A novel end-to-end deep learning solution for coronary artery segmentation from CCTA. Med Phys. 2022 Nov 1;49(11):6945\u0026ndash;59. \u003c/li\u003e\n\u003cli\u003eAlgarni M, Al-Rezqi A, Saeed F, Alsaeedi A, Ghabban F. Multi-constraints based deep learning model for automated segmentation and diagnosis of coronary artery disease in X-ray angiographic images. PeerJ Comput Sci. 2022;8. \u003c/li\u003e\n\u003cli\u003eShen Y, Chen Z, Tong J, Jiang N, Ning Y. DBCU-Net: deep learning approach for segmentation of coronary angiography images. International Journal of Cardiovascular Imaging. 2023 Aug 1;39(8):1571\u0026ndash;9. \u003c/li\u003e\n\u003cli\u003eSerrano-Anton B, Otero-Cacho A, Lopez-Otero D, Diaz-Fernandez B, Bastos-Fernandez M, Perez-Munuzuri V, et al. Coronary Artery Segmentation Based on Transfer Learning and UNet Architecture on Computed Tomography Coronary Angiography Images. IEEE Access. 2023;11:75484\u0026ndash;96. \u003c/li\u003e\n\u003cli\u003eKaba Ş, Haci H, Isin A, Ilhan A, Conkbayir C. The Application of Deep Learning for the Segmentation and Classification of Coronary Arteries. Diagnostics. 2023 Jul 1;13(13). \u003c/li\u003e\n\u003cli\u003eLiu X, Wang X, Chen D, Zhang H. Automatic Quantitative Coronary Analysis Based on Deep Learning. Applied Sciences (Switzerland). 2023 Mar 1;13(5). \u003c/li\u003e\n\u003cli\u003eZhang X, Zhang B, Zhang F. Stenosis Detection and Quantification of Coronary Artery Using Machine Learning and Deep Learning. Vol. 75, Angiology. SAGE Publications Inc.; 2024. p. 405\u0026ndash;16. \u003c/li\u003e\n\u003cli\u003eLi S, Fan Y. Coronary Artery Segmentation in X-ray Angiography Based on Deep Learning Approach. \u003c/li\u003e\n\u003cli\u003eNannini G, Saitta S, Baggiano A, Maragna R, Mushtaq S, Pontone G, et al. A fully automated deep learning approach for coronary artery segmentation and comprehensive characterization. APL Bioeng. 2024 Mar 1;8(1). \u003c/li\u003e\n\u003cli\u003eJiang W, Li Y, Jia Y, Feng Y, Yi Z, Wang J, et al. Segmentation of coronary artery based on discriminative frequency learning and coronary-geometric refinement. Comput Biol Med. 2024 Oct 1;181. \u003c/li\u003e\n\u003cli\u003eMuthusamy CD, Murugesh R. Integrated deep learning approach for automatic coronary artery segmentation and classification on computed tomographic coronary angiography. Network Modeling Analysis in Health Informatics and Bioinformatics. 2024 Dec 1;13(1). \u003c/li\u003e\n\u003cli\u003eWang L, Zhang X, Tian C, Chen S, Deng Y, Liao X, et al. PlaqueNet: deep learning enabled coronary artery plaque segmentation from coronary computed tomography angiography. Vis Comput Ind Biomed Art. 2024 Dec 1;7(1). \u003c/li\u003e\n\u003cli\u003eMahendiran T, Thanou D, Senouf O, Jamaa Y, Fournier S, De Bruyne B, et al. AngioPy Segmentation: An open-source, user-guided deep learning tool for coronary artery segmentation. Int J Cardiol. 2025 Jan 1;418. \u003c/li\u003e\n\u003cli\u003eMorshedul Abedin AJ, Sarmun R, Mushtak A, Bin Mohamed Ali MS, Hasan A, Suganthan PN, et al. Enhanced coronary artery segmentation and stenosis detection: Leveraging novel deep learning techniques. Biomed Signal Process Control. 2025 Nov 1;109. \u003c/li\u003e\n\u003cli\u003eShrivastava P, Kashikar S, Parihar PH, Kasat P, Bhangale P, Shrivastava P. A systematic review on deep learning-enabled coronary CT angiography for plaque and stenosis quantification and cardiac risk prediction. Eur J Radiol Open. 2025 Jun 1;14. \u003c/li\u003e\n\u003cli\u003eAlskaf E, Dutta U, Scannell CM, Chiribiri A. Deep learning applications in coronary anatomy imaging: a systematic review and meta-analysis. Available from: https://jmai.amegroups.com/article/\u003c/li\u003e\n\u003cli\u003eGharleghi R, Chen N, Sowmya A, Beier S. Towards automated coronary artery segmentation: A systematic review. Vol. 225, Computer Methods and Programs in Biomedicine. Elsevier Ireland Ltd; 2022. \u003c/li\u003e\n\u003cli\u003eTu L, Deng Y, Chen Y, Luo Y. Accuracy of deep learning in the differential diagnosis of coronary artery stenosis: a systematic review and meta-analysis. BMC Med Imaging. 2024 Dec 1;24(1). \u003c/li\u003e\n\u003cli\u003eAlnasser TN, Abdulaal L, Maiter A, Sharkey M, Dwivedi K, Salehi M, et al. Advancements in cardiac structures segmentation: a comprehensive systematic review of deep learning in CT imaging. Vol. 11, Frontiers in Cardiovascular Medicine. Frontiers Media SA; 2024. \u003c/li\u003e\n\u003cli\u003eWang L, Liang D, Yin X, Qiu J, Yang Z, Xing J, et al. Coronary artery segmentation in angiographic videos utilizing spatial-temporal information. BMC Med Imaging. 2020 Sep 24;20(1):110. \u003c/li\u003e\n\u003cli\u003eMaxim Popov AANZAAASTA et al. ARCADE: Automatic Region-based Coronary Artery Disease diagnostics using x-ray angiography imagEs Dataset. 2023; \u003c/li\u003e\n\u003cli\u003ePopov M, Amanturdieva A, Zhaksylyk N, Alkanov A, Saniyazbekov A, Aimyshev T, et al. Dataset for Automatic Region-based Coronary Artery Disease Diagnostics Using X-Ray Angiography Images. Sci Data. 2024;11(1):1\u0026ndash;9. \u003c/li\u003e\n\u003cli\u003eRonneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2015;9351:234\u0026ndash;41. \u003c/li\u003e\n\u003cli\u003eZhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation BT - Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Miccai. 2018;11045(2018):3\u0026ndash;11. \u003c/li\u003e\n\u003cli\u003eXie E, Wang W, Yu Z, Anandkumar A, Alvarez JM, Luo P. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. Adv Neural Inf Process Syst. 2021;15(NeurIPS):12077\u0026ndash;90. \u003c/li\u003e\n\u003cli\u003eKaba Ş, Haci H, Isin A, Ilhan A, Conkbayir C. The Application of Deep Learning for the Segmentation and Classification of Coronary Arteries. Diagnostics. 2023 Jul 1;13(13). \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Coronary artery disease, Coronary angiography, Deep learning, Image segmentation, Inference time","lastPublishedDoi":"10.21203/rs.3.rs-7467606/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7467606/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eCoronary artery disease (CAD) is one of the leading causes of cardiovascular mortality worldwide. Accurate segmentation of coronary arteries from X-ray coronary angiography (XCA) images is crucial for assessing vessel morphology and stenosis, thereby supporting computer-aided diagnosis and guiding interventional treatment decisions. Although recent studies have primarily focused on enhancing segmentation accuracy using deep learning models, limited attention has been given to evaluating their inference time \u0026mdash; a factor that is equally important for clinical deployment and real-time decision support.\u003c/p\u003e\u003ch2\u003eObjective\u003c/h2\u003e\u003cp\u003eThis study compares both segmentation performance and inference time of U-Net, U-Net++, and SegFormer on the ARCADE XCA dataset (stenosis and SYNTAX subsets).\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eAll XCA images were resized to 256 \u0026times; 256 pixels, normalized, and augmented prior to training. The U-Net and U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;architectures were implemented as convolutional encoder\u0026ndash;decoder networks with skip connections, whereas SegFormer employed a hierarchical Transformer-based encoder coupled with a lightweight MLP decoder. All models were trained for 100 epoch using cross-entropy loss with class-balancing weights. Performance was evaluated in terms of segmentation accuracy, dice score, and per-image inference time.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eOn the stenosis subset, U-Net and U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;achieved the highest training accuracy (99.82%), while SegFormer attained a slightly lower accuracy (99.15%) but delivered the fastest inference time (0.05 s per image). On the SYNTAX subset, U-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;obtained the best training accuracy (98.13%), followed closely by U-Net (98.04%) and SegFormer (97.00%). Despite its lower accuracy, SegFormer consistently demonstrated superior efficiency, achieving the shortest inference time (0.18 s per image).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eU-Net\u0026thinsp;+\u0026thinsp;+\u0026thinsp;demonstrated the highest segmentation accuracy, SegFormer provided the most significant runtime advantage, and U-Net achieved a balanced trade-off between the two. Taken together, these findings suggest that model selection should be informed by the specific priorities of clinical deployment, whether accuracy, inference speed, or a compromise between both is most critical.\u003c/p\u003e","manuscriptTitle":"Comparative Analysis of Deep Learning Models for Coronary Artery Segmentation: Performance and Inference Time Evaluation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-03 16:32:30","doi":"10.21203/rs.3.rs-7467606/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2025-10-15T14:53:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"1921005748703650992097206330865511688","date":"2025-09-29T17:14:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"149909070490053173721758417610367554657","date":"2025-09-24T11:37:45+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-22T10:53:55+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-08-28T10:31:13+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-28T08:51:47+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-28T08:49:50+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-08-27T04:04:38+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5bcf4561-e119-451a-9817-5f8d2f2e0871","owner":[],"postedDate":"October 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-10-03T16:32:30+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-03 16:32:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7467606","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7467606","identity":"rs-7467606","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}