Source-linked AI summary
ImageCAS: A Large-Scale Dataset and Benchmark for Coronary Artery Segmentation based on Computed Tomography Angiography Images
An Zeng, Chunbiao Wu, Meiping Huang, Jian Zhuang, Shanshan Bi, Dan Pan, Najeeb Ullah, Kaleem Nawaz Khan, Tianchen Wang, Yiyu Shi, Xiaomeng Li, Guisen Lin, Xiaowei Xu
TL;DR
Coronary artery segmentation on CTA is important but difficult to compare across studies because prior work often uses small, private, or inconsistent datasets. This paper introduces the 1,000-patient ImageCAS dataset and a benchmark with existing methods plus a multi-scale patch-fusion, two-stage baseline. The baseline achieves the optimal benchmark performance, including a Dice score of 82.96%, although performance still has room for improvement in clinical diagnosis and stenosis quantification.
Problem
Prior coronary artery segmentation studies often rely on small or private datasets, unpublished code, and inconsistent comparisons, limiting fair evaluation.
Method
The paper constructs a public 1,000-patient CTA dataset, implements typical existing methods in a benchmark, and proposes a multi-scale patch-fusion, two-stage baseline.
Results
The proposed baseline achieves the optimal benchmark performance with a Dice score of 82.96%.
Takeaways & Limitations
ImageCAS provides a public dataset and benchmark for more uniform comparison and further development of coronary artery segmentation methods.
Takeaways & Limitations
The paper reports that benchmark implementations may miss critical details from prior studies, and performance remains insufficient for fully accurate clinical diagnosis and stenosis quantification.
Abstract
from arXiv · showhide
Cardiovascular disease (CVD) accounts for about half of non-communicable diseases. Vessel stenosis in the coronary artery is considered to be the major risk of CVD. Computed tomography angiography (CTA) is one of the widely used noninvasive imaging modalities in coronary artery diagnosis due to its superior image resolution. Clinically, segmentation of coronary arteries is essential for the diagnosis and quantification of coronary artery disease. Recently, a variety of works have been proposed to address this problem. However, on one hand, most works rely on in-house datasets, and only a few works published their datasets to the public which only contain tens of images. On the other hand, their source code have not been published, and most follow-up works have not made comparison with existing works, which makes it difficult to judge the effectiveness of the methods and hinders the further exploration of this challenging yet critical problem in the community. In this paper, we propose a large-scale dataset for coronary artery segmentation on CTA images. In addition, we have implemented a benchmark in which we have tried our best to implement several typical existing methods. Furthermore, we propose a strong baseline method which combines multi-scale patch fusion and two-stage processing to extract the details of vessels. Comprehensive experiments show that the proposed method achieves better performance than existing works on the proposed large-scale dataset. The benchmark and the dataset are published at https://github.com/XiaoweiXu/ImageCAS-A-Large-Scale-Dataset-and-Benchmark-for-Coronary-Artery-Segmentation-based-on-CT.
1 INTRODUCTION
Coronary artery segmentation supports CTA-based diagnosis, but anatomical complexity and inconsistent prior datasets and evaluations make fair method comparison difficult. The paper addresses this gap with a 1,000-patient dataset, a reproducible benchmark, and a strong baseline.
- Motivation: Coronary artery segmentation is important for diagnosing and quantifying coronary artery disease from CTA images.CTA is noninvasive and provides high-resolution 3D imaging, while manual segmentation is time-consuming and error-prone.
- Challenges: Anatomical variation, imaging artifacts, complex branching, and small vessel area make coronary artery segmentation challenging.Coronary arteries may be surrounded by fat or embedded in heart muscle, while bifurcations and noisy scans complicate segmentation.
- Research gap: Prior studies often used private or small datasets and inconsistent annotations, metrics, and comparisons, limiting fair evaluation.Some studies compared results across different datasets or included the initial aorta in annotations, producing incomparable Dice scores.
- Contributions: The paper introduces a publicly available dataset containing 1,000 patients and a benchmark implementing several existing coronary artery segmentation methods.The dataset and benchmark are intended to support uniform investigation and fair comparisons.
- Contributions: The proposed baseline combines multi-scale patch fusion with two-stage processing and outperforms existing state-of-the-art methods.The method is designed to retain global context while recovering local vessel details.
2 RELATED WORK
The paper reviews traditional and deep-learning approaches for coronary artery segmentation and identifies persistent weaknesses in public data and benchmarking practice. It motivates ImageCAS as a larger, public evaluation setting with implemented comparison methods.
- 2. RELATED WORK: Coronary artery segmentation research includes traditional machine-learning methods and deep-learning trends spanning pixel, direct, patch, tree, and graph representations.Traditional approaches use hand-crafted features and structural priors, while deep methods increasingly model image regions and vessel topology.
- 2. RELATED WORK: Direct segmentation processes whole images but commonly requires downsampling, whereas patch-based methods refine downsampled regions to recover local details.Patch-based approaches address computational constraints and information loss associated with reduced-resolution inputs.
- 2.3 Benchmark and Datasets: Published studies frequently rely on small or proprietary datasets and use inconsistent datasets, metrics, annotations, or comparison baselines.These inconsistencies include evaluation on only 42 public CTA images and annotation schemes that include the initial aorta.
- 2.3 Benchmark and Datasets: The paper argues that a public benchmark is needed because existing datasets are small and prior implementations and comparisons are often incomplete.Common datasets cited for deep learning contain only 8 and 18 training images, while source implementations and experimental details are frequently unavailable.
- 2.3 Benchmark and Datasets: ImageCAS provides a larger public dataset and implements typical existing methods to support more uniform evaluation.The benchmark includes direct, patch-based, tree-based, and graph-based segmentation approaches.
3 IMAGECAS DATASET
ImageCAS comprises 3D CTA scans from 1,000 patients with coronary artery annotations produced through radiologist review and consensus. The dataset includes detailed volumetric imaging specifications and labels based on coronary artery anatomy.
- 3 IMAGECAS DATASET: ImageCAS contains 3D CTA images from 1,000 patients acquired using a Siemens 128-slice dual-source scanner.The scans use selected cardiac phases and have sizes of 512×512×(206−275) voxels with planar resolution of 0.29-0.43 mm2.
- 3 IMAGECAS DATASET: The benchmark methods include direct, patch-based, tree-based, and graph-based segmentation approaches.These method categories are summarized in the benchmark overview.
- 3 IMAGECAS DATASET: Each image’s left and right coronary arteries were independently labeled by two radiologists, with discrepancies resolved by a third radiologist through consensus.This annotation process provides crossvalidation before final labels are determined.
- 3 IMAGECAS DATASET: The labels cover coronary artery branches listed under the AHA naming convention.The listed anatomy includes the left main, left anterior descending, left circumflex, right coronary, diagonal, obtuse marginal, ramus intermedius, posterior descending, and acute marginal arteries.
4 BENCHMARK
The benchmark implements representative direct, patch-based, tree-based, graph-based, and proposed baseline methods for coronary artery segmentation. The baseline combines coarse and patch segmentation to balance global context, local detail, and computational feasibility.
- Benchmark methods: The benchmark evaluates direct, patch-based, tree-based, graph-based, and baseline coronary artery segmentation methods.The implementations are based on representative existing approaches, despite unavailable source code and incomplete experimental details for some methods.
- Direct segmentation: Direct segmentation resizes CTA images, applies an FCN with attention gates, and upsamples the prediction to the original resolution.This approach is simple but can lose local detail because large CTA images must be downsampled for processing.
- Patch-based segmentation: Patch-based segmentation first identifies a region of interest, then processes localized image patches to address computational constraints and missing details from downsampling.The benchmark considers a dual-CNN framework using a 3D U-net for region extraction and subsequent patch processing.
- Baseline method: The baseline combines coarse segmentation with multi-scale patch segmentation to preserve global structure while recovering vessel details.Coarse segmentation uses a 3D U-net, while patch segmentation uses skeleton-centered cubic patches processed by 3D U-net++.
- Baseline method: The baseline uses dilated vessel targets and further test-time dilation to encourage coverage and connectivity before skeleton-guided patch extraction.The weighted loss uses α=0.01, biasing outputs toward oversized vessels during training.
5 EXPERIMENTS AND DISCUSSION
The experiments establish a common evaluation setup before comparing each method under specific configurations and then analyzing optimally configured methods.
- The experiments first define a common setup, then evaluate method-specific configurations, and finally compare all methods using their optimal configurations.
5.1 Experiment setup
The study uses PyTorch and DGL on an RTX 3090 GPU, with method-dependent losses and standardized training settings. Direct-segmentation configuration effects are visualized through input size, attention gates, and channel count.
- Implementation: Experiments were implemented in PyTorch and DGL on an Nvidia RTX 3090 GPU with 24G memory.
- Direct-segmentation analysis: Figure 4 examines direct segmentation across input size, attention-gate use, and channel count, with significance markers ranging from not significant to p<0.0001.
- Visual analysis: Figure 5 provides visual examples of direct-segmentation results, including failure cases.
- Training setup: The study trains networks for 30 epochs with Adam at learning rate 0.002, while batch size varies with input size because of GPU memory limits.Batch sizes are 8, 2, and 1 for the listed input sizes 128 × 128 × 128, 256×256×128, and 512×512×256, respectively.
5.2 Configuration discussion
Configuration studies show that resolution, attention, channel count, patch context, and preprocessing choices materially affect coronary artery segmentation, while failure analysis identifies limits in patch, tree, and graph pipelines.
- 5.2.1 Direct segmentation: 512×512×256 input improved Dice by 7.38% and 12.32% over smaller inputs, while attention and 12 channels added 1.34% and 2.13%, respectively.The comparisons were statistically significant at p<0.0001.
- 5.2.2 Patch based segmentation: Patch size mattered more than Frangi features: the Frangi channel changed Dice by only 0.01%, whereas larger patches improved context and disabling flipping and rotation improved Dice by 2.63% and 2.73%.The authors attribute harmful augmentation to anatomically unrealistic vessel orientations and surrounding anatomy.
- 5.2.3 Tree data based segmentation and graph based based segmentation: High-resolution input improved graph segmentation by 2.95% Dice but changed tree segmentation by only 0.12%, while the two tree models differed by 0.06%.In tree segmentation, 16 × 16 × 4 patches exceeded 16 × 16 × 8 patches by 1.38% Dice.
- Failure analysis: Pre-segmentation errors can omit coronary arteries from subsequent tree or graph construction, and patch post-processing can remove true vessels while suppressing similar structures.Failed cases include missing vessels, bone tissue, and small-vessel artifacts.
- 5.2.4 Baseline method: 82.96% Dice was achieved by ensembling coarse and patch segmentation, outperforming the individual base classifiers at 77.80%, 82.27%, and 82.70%.The baseline study also examined dilation and patch sizes as intermediate design choices.
5.3 Benchmark comparison
The benchmark compares five coronary artery segmentation methods across quantitative and visual evaluations. The proposed baseline performs best overall, while implementation and dataset differences limit direct comparison with prior work.
- The proposed baseline achieves optimal performance on all benchmark metrics.
- Table 4 reports the benchmark’s method performance using each implementation’s optimal configuration.
- Cross-paper comparisons remain difficult because prior studies use different datasets, metrics, annotations, and implementation details.
- Direct, patch, tree-based, and graph-based methods show differing strengths across cases, particularly for thin or low-contrast vessels.
- Visual cases show the baseline more reliably recognizes vessels, especially low-contrast vessels in difficult images.Its multi-scale patch fusion combines coarse and patch-level features to capture context information.
5.4 Discussion
The benchmark establishes a strong baseline but also identifies dataset bias, scanner restriction, missing detailed labels, and opportunities for stronger architectures and topology-aware evaluation.
- The benchmark baseline achieves optimal Dice, HD, and AHD performance among the implemented methods.
- The dataset is limited by collection at one center and acquisition with a single Siemens 128-slice dual-source scanner.These choices introduce biased distributions and may worsen dataset bias.
- The dataset does not provide detailed labels separating coronary artery subclasses such as the left main and left anterior descending arteries.
- Future work could test advanced segmentation networks and metrics that preserve coronary-vessel topology or connectivity.
6 CONCLUSION
The paper introduces a public CTA dataset and benchmark for coronary artery segmentation, including a strong baseline and implementations of typical methods. The baseline reaches a Dice score of 82.96%, while the dataset and benchmark are intended to stimulate further research.
- The proposed baseline achieves the optimal benchmark performance with a Dice score of 82.96%.
- The work provides a benchmark dataset and implemented comparison methods for coronary artery segmentation on CTA images.
- The reported performance still has room for improvement for accurate diagnosis and stenosis quantification in clinical practice.
- The benchmark and dataset are publicly published to stimulate further community research.
7 ETHICAL AND INFORMATION GOVERNANCE AP-
The study received ethics approval and deidentified retrospective CT data before analysis.
- The retrospective data collection received ethics approval under Protocol No. 2019324H.
- Patient-sensitive information was deidentified during conversion of CT files to NIfTI format.