Semantic segmentation of coronary arteries in computed tomography angiograph: A multi-center, multi-vendor and multi disease study

doi:10.21203/rs.3.rs-4688114/v1

Download PDF

Article

Semantic segmentation of coronary arteries in computed tomography angiograph: A multi-center, multi-vendor and multi disease study

https://doi.org/10.21203/rs.3.rs-4688114/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Coronary artery disease (CAD) is one of the primary causes of death worldwide accountings for almost 25% of all deaths worldwide each year. Accurate diagnosis and assessment of CAD and stenosis require the semantic segmentation of the coronary arteries and aorta in coronary computed tomography angiography (CCTA) images, which is of great significance. Nonetheless, producing semantic segmentations is particularly challenging due to the anatomical similarity between various surrounding areas, the branching of the arteries which includes numerous bifurcations, and the presence of smaller vessels that add to the complexity. Reducing the resolution of a 3D image to fit within the constraints of available GPU memory can result in a loss of detail, which is often undesirable. Instead, using patches of the image as input can help mitigate this issue. In this study, we propose a novel semantic segmentation method based on the 3D U-Net that uses three different datasets consisting of multi-center, multi-vendor, and multi-continent data CCTA images. We achieved accuracies of 91.2%, 93.9%, and 97.3%, respectively. The outcomes of the study demonstrate the efficacy of the method for accurately segmenting the aorta and/or the coronary arteries.

coronary arteries

semantic segmentation

CCTA

deep learning

3D U-net

In a world where coronary artery disease (CAD) diagnosis claim millions of lives each year [1], early detection and treatment are of paramount importance [2]. The health of the heart is heavily dependent on the patency of the coronary arteries, which feed blood to its muscles. The deposition of cholesterol and fats in these arteries can cause plaque development, resulting in coronary artery stenosis and limiting blood flow, compromising myocardial oxygenation [2]. As a result, accurate CAD diagnosis and risk assessment are critical to minimize its impact.

Coronary computed tomography angiography (CCTA) plays a crucial role in this effort as it is a non-invasive technique that offers complex imaging capabilities and allows for reliable analysis and prognostic insights for diagnosing CAD. Particularly in the case of coronary artery disease (CAD) diagnosis, CCTA is a reference modality. It allows the assessment of the arteries and the identification of abnormalities through segmentation [3].

Coronary artery segmentation in CCTA images is a particularly challenging task [4]. Firstly, there are several branches in the coronary circulation that vary in thickness, and some of these branches are too thin to be precisely segmented due to their complex structure. Individual differences may also be observed in the coronary artery tree morphology. Secondly, there are other kinds of vascular structures that resemble the coronary arteries around the heart and are sometimes misidentified as such. Thirdly, the segmentation techniques must take into consideration the fact that the coronary arteries represent a very small portion of the heart's overall tissue [5]. Furthermore, the gray information of the coronary artery in the CCTA image is similar to that of the bone and heart region, although it differs significantly from the background. The clarity of CCTA images varies significantly due to a number of factors, including varying patients’ anatomy, unequal contrast agents, complicated coronary vascular architecture, similar morphology between different types of arteries. Last but not least, the heart rate, and data reconstruction method affect the quality of the images, resulting in low-quality images [4]. To address these issues, segmentation and feature extraction algorithms for CCTA images have to satisfy specific requirements. First, vascular plaque characteristics must be extracted in a rich and accurate manner that can detect small changes. Second, the algorithms must exhibit the ability to deal with any challenges caused by a variety of adverse circumstances and challenging situations. Finally, given the time limits in practical applications, these algorithms must be simple to implement, fast to compute, and have a low calculation complexity.

The semantic segmentation of coronary arteries is extremely important for clinical interventional cardiology practice. Determining the percent stenosis of coronary artery accurately requires real-time guidance for CAD clinical decisions [6]. Radiologists typically examine the entire vascular tree based on its morphology and position. In light of these requirements, this study concentrates on CCTA images, examining semantic segmentation techniques in medical image processing, which it forms a key part of the process of image segmentation. The objective is to find an algorithm to align with clinical application needs and deal with different data parameters, with a particular focus on image semantic segmentation based on a 3D U-net model.

Each year, a large variety of advanced techniques are developed for coronary artery segmentation of various medical images, including CCTA, X-ray angiography (CAG) and fluoroscopy angiography. These methods fall into two categories: segmentation approaches and semantic segmentation approaches. The semantic segmentation algorithms have demonstrated cutting-edge performance in coronary artery segmentation in fluoroscopic images [7]. The main step to segmenting the coronary artery in this approach is to extract the vascular tree and the major vessel, then classify each individual coronary artery according to various classes by extracting underlying features of the vessel segments (arterial topology, position, and pixel features). In our knowledge, this is the only work [7] in semantic segmentation in this field, and it is based on classifying the vessels according to some parameters, not on annotated data using fluoroscopy angiography.

In this paper, we aim to develop a precise method for the semantic segmentation of coronary arteries and aorta, designed to be effective across acquisition devices and institutions. To this end, we develop and validate a deep learning-based approach using, three diverse datasets sourced from three different centers countries and three continents. This allows us to capture interinstitutional variability, which is influenced significantly by the machines used, acquisition protocols, and patient characteristics.

Moreover, we prepared annotations for these datasets, which accurate annotations by medical experts can be rare and costly in terms of time. This is the first study that focuses on coronary artery semantic segmentation in CCTA images using multi-continent, multi-center and multi-vendor, providing a relevant multi-country, multi-continent dataset. The primary contributions of this work are:

Annotation of three multi-center, multi-continent data sets for sematic segmentation models
Patches based segmentation approaches to improve semantic segmentation accuracy
Sematic segmentation using 3D Unet for multi-center and multi-vendor datasets

The remainder of this paper is organized as follows: Section two devoted to related work. Section three to the materials and methods used in this paper. In section four, we present the obtained results. Section five we present the discusses of the results, and the paper concludes with section six, which offers the main conclusions.

In recent times, the state-of-art architectures rely on full convolutional neural network (FCN)[8] and U-Net [9] in the field of coronary artery segmentation. These neural network architectures are capable of extracting crucial information from spatial features, which is essential for segmenting objects. However, most modern semantic segmentation models in this field have been utilised are in coronary angiography modality such as [7]. Recently, Zeng et al. [10] conducted an exhaustive literature review of existing research in the field of coronary artery segmentation, where they addressed the issue of a lack of publicly available databases, making benchmarking of segmentation algorithms extremely difficult. To address this issue, a large-scale dataset of 1000 scans was made publicly available, along with a baseline approach for segmentation. They additionally implemented numerous common existing methods. The proposed baseline technique performed better than previous approaches on the presented dataset. Song et al. [5] proposed a deep convolutional neural network (CNN)-based automatic segmentation approach named 3D Feature-Efficient-Fusion-UNet. First, non-coronary-artery slices are screened out using a 2D DenseNet classification network. Then, a 3D-UNet-based coronary artery segmentation network is used, consisting of a 3D-FEF Dense Block for encoding and 3D-Residual Blocks for decoding. Segmentation results showed that the proposed 3D-FEF-UNet achieved a value of Dice-Score coefficient (DSC) (82%). Xin et al. [6] proposed an automatic method for coronary plaque detection and classification that takes advantage of both (CNN) and radiomics. A modified Mask-Region based CNN model is utilized to segment the coronary arteries, identify coronary plaque candidates, and extract the image patch with great computational efficiency. He et al. [11] proposed a new semi-supervised framework for whole coronary arterial tree segmentation in invasive coronary angiography. A mean Teacher (MT) model was used as the main framework with Nested U-Net (UNet++) as its backbone. The proposed method yields a Dice score of 81.66%. Li et al. [12] implemented a deep learning model for automatic CDA diagnosis consisting of two subtasks. The first of them is a segmentation task that uses U-Net to autonomously segment and extract the coronary artery. The second one is an identification of the coronary arteries task that makes use of the 3DNet technique. In coronary artery automatic-segmentation, the model attained a DSC value of 0.771. Pan et al. [13] proposed a 3D Dense-U-Net combined with the focal loss to tackle the class imbalance between the background and the coronary artery. The results showed a DSC of 0.9691, which is higher than that of the 3D UNet. Gu & Cai. [14] proposed a two-stage strategy for segmenting the aorta and coronary arteries that combines the benefits of both 2D and 3D CNNs. In the first stage, a 2D CNN is employed to segment the aorta and coronary arteries simultaneously, slice by slice. Then, in the second stage, a 3D CNN is used to refine the segmentation of the coronary arteries by retrieving inter-slice information that was missed in the previous stage. To validate the efficacy of the method, they compared it to other approaches that used pure 2D CNNs, 3D CNNs, and hybrid 2D-3D CNNs. The average DSC values for the aorta and coronary arteries are 97.54% and 86.62%, respectively. To solve the issue of limited GPU memory, Tian et al. [15] proposed a distinct image segmentation framework that matches deep learning with digital image processing approaches. They applied the V-Net and a cascaded V-shaped network to fully utilize the image context information. The morphology operation and the fundamental gradient are then used as a condition to terminate the region growing search, making the coronary segmentation edge more realistic. Compared with other methods as 3D-Unet and AGnet, the proposed method achieves higher segmentation accuracy (DSC 93%), it can accurately identify small coronary arteries, and the borders of segmented coronary arteries are smooth and filled. Zhu et al. [16] proposed a Pyramid Scene Parsing Network (PSP-Net) based multi-scale CNN model. This algorithm first down-samples the pre-processed image to obtain images of multiple scales, and then sends them to the CNN. The features of different scales are selected and fused in the fully connected layer, and finally the features are used to segment images. It should be noted that this study compares traditional image segmentation methods and methods using deep learning. The accuracy of the proposed PSPNet is 0.95. Cheung et al. [17] employed a fully automated 2D-Unet model for segmenting the aorta and coronary arteries. The model achieved DSC accuracy of 91.20% and 88.80% for aorta and coronary artery respectively. When the aorta and coronary arteries are segmented, the model performance decreases. On the other hand, segmenting the coronary arteries alone yields better accuracy. Zhao et al. [7] proposed a novel graph-based hybrid deep learning method for the extraction of individual coronary arteries. Firstly, the entire vascular tree and major vessels are extracted by Feature pyramid UNet + + network. Then, by generating the vascular tree centerline to discover the critical points and converting it into a vessel graph, each blood vessel segment has been expressed by an edge in the graph. Lastly, the semantic segmentation findings are generated after classifying the vessel segments according to various classes, by extracting underlying features of the vessel segments (arterial topology, position and pixel features). Lei et al. [18] proposed a 3D attention Fully Convolution Network (FCN), in which the FCN was used for segmentation and the deep attention method was incorporated in order to highlight semantic information. The method yielded a DSC value of 83%. [19] Xian et al. applied four deep neural networks (DNNs) for main vessel segmentation of X-ray angiography. Based on the UNet model, the method replaced its backbone by three widely image classification networks (ResNet, DenseNet, and Residual Attention Network (ResAttNet). In terms of precision, the models with changed backbones outperform the UNet. ResAttNet specifically attained a DSC precision of 0.919. Baskaran et al. [20] used a derived U-net model for the identification of multiple cardiovascuar structures. Huang et al. [21] applied an automated wall segmentation by detecting boundaries of the coronary artery (inner & outer boundaries) that concentrates on the continuity and connection of boundaries. Gu et al. [22] proposed a 3D global feature embedded network (GFE-Net) with active contour loss for coronary artery segmentation. The global feature combines the comprehensive information of multi-level features with the goal of precisely segmenting targets and defined boundaries using semantic information and detailed attributes. An improved learning active contour model is further employed to obtain refined segmentation results. The average DSC is 0.9143%. Chen et al. [23] proposed a 3D multi-channel U-Net architecture, in which the tubular structure of coronary arteries is highlighted by the inclusion of a vesselness map in the U-Net input, helping the network to learn more effectively and achieved a value of 80% of DSC accuracy. Shen et al. [24] utilized a deep learning and level-set approach framework to segment both the aorta and coronary arteries. The 3D semantic properties of coronary arteries are learned using a 3D fully convolutional network (FCN), and an attention gate mechanism is also incorporated into it to suppress feature activation in irrelevant regions and improve the accuracy of the model. The level set function is integrated with 3D FCN to refine and optimize the results. The mean DSC accuracy is 90%. Yang et al. [25] developed an approach for major vessel segmentation by integrating U-Net architecture and deep convolutional networks. This involved replacing the popular image classification networks ResNet101, DenseNet121, and InceptionResNet-v2 for the U-Net backbone. The hybrid networks outperformed simple U-Net in terms of recall, precision. P. Mirunalini et al. [26] proposed a combined deep learning model involving identification of the existence of the coronary arteries in the 2D slices. The slices which contain the coronary arteries are detected with Recurrent Convolutional Neural Network R-CNN, then segmented by a 2D residual U-Net architecture. The method attained an intersection over union (IoU) of 84%. Yang et al. [27] applied five fully convolutional networks by combining simple U-Net, VGG16, ResNet, DenseNet, and InceptionResNetv2 encoders with U-Net decoder. DenseNet121 attained the greatest DSC average, 91.9%. Wolterink et al. [28] used an approach to extract coronary artery centerlines utilizing a 3D dilated CNN trained to predict the majority of typical direction and radius of an artery at any given position depending on a local image patch in the images. [29] Wolterink et al. directly optimized the spatial placement of vertices in a tubular surface mesh, which segments the coronary artery lumen using graph convolutional networks (GCNs). Accurate and efficient extraction of the coronary artery surface grid is achieved. 74% is the average DSC. [30] Huang et al. employed a 3D U-Net to segment both coronary arteries with and without centerline. The results showed a DSC between 71% and 78%. Nasr-Esfahani et al. [31] proposed a method for detecting vessel regions which is based on two convolutional neural networks (CNN). The first CNN uses local and global features to form a vessel probability map. The second CNN, includes both the edge map and the vessel probability map to find an enhanced vessel probability map as well as to correct the outcome of the first stage. The reported accuracy was 97.93%. Kjerland [32] used two neural networks which are trained on aortic and coronary segmentation respectively. In some images, the method successfully segments the entire coronary artery tree, but fails to segment all branches in the rest of the images. The DSC accuracy was reported to be between 75% and 78%.

A. Dataset

First dataset

The first dataset was obtained at a private clinic “Medical Imaging Centre Dr Lekhal” in Algeria. The CCTA scans were acquired using a Revolution CT scanner from (GE Healthcaree), using retrospective acquisition synchronised with the ECG. Reconstruction was performed over the different phases of the R-R interval of the cardiac cycle. We collected 90 scans, 48 of which were reconstructed at phase 75% of the R-R interval. Patients were aged between 28 and 89 years. All scans were acquired with a tube voltage of 100 kVp. Images were reconstructed with 0.48 × 0.48-pixel pitch and 0 slice thickness. The process for annotating the coronary arteries and the descending aorta is as follows: The radiologist uses the volume reconstructed by the semi-automatic method, which is based on the detection of the axis of the coronary artery on the AW workstation (GE Healthcare product), to select the most appropriate phase according to the heart rate of the patient. Subsequently, the final volume is saved and provided to the patient along with the medical report. The CCTA images and the corresponding masks can then be downloaded from the station in Dicom format.

This integration of expert radiological analysis with advanced software solutions significantly enhances the accuracy and speed of coronary artery annotation, minimizing errors compared to manual labeling by technologists alone, especially in areas where the arteries are too small to be seen by the human eye.

In this dataset, each scan consisted of 3D images of the heart with a resolution of 512 × 512 × 256 voxels. Initially, the images were in DICOM format, but where then converted to NIFTI.

The scans are splitted into a training set of 50 scans and a validation set of 18 scans.

No individual consent was necessary according to national guidelines since no identifiable data was collected

Second dataset

Our second dataset sourced from the Medical Imaging and International Clinical Research Center (ICRC) in Czech Republic. We provide a detailed descriptions of the CT imaging protocols using the Philips Brilliance iCT 256 scanners. For patients with a heart rate above 65 (111 patients), a retrospective ECG-gated protocol was employed. The protocol details include a slice thickness/increment of 0.9/0.45 mm, a radiation dose of 600 mAs, and kV settings based on weight (80 kV for 9 patients, 100 kV for 47 patients, and 120 kV for 45 patients). The imaging parameters included a 512 matrix, 128x0.625 mm collimation, iDose 6, a field of view (FOV) of 250 mm, and reconstructed phases typically at 0%, 25%, 50%, and 75%. For patients with a heart rate below 65 (17 patients), a prospective ECG-gated protocol was used, with similar thickness/increment settings, but a reduced radiation dose of 200 mAs and kV settings of 80 kV for 4 patients, 100 kV for 8 patients, and 120 kV for 5 patients, maintaining other parameters consistent with the retrospective protocol, and usually reconstructing at 78%. Additionally, a specialized retrospective ECG-gated protocol with a larger FOV of 350 mm was applied to 10 patients for more extensive aorta imaging, with a 0.9/0.45 mm thickness/increment, 600 mAs, 100 kV, a 512 matrix, 96x0.625 mm collimation, iDose 5, and a reconstruction phase at 75%. Lastly, a higher resolution retrospective ECG-gated protocol was used for 3 patients with a 0.67/0.335 mm thickness/increment, 500 mAs, 100 kV, a 512 matrix, 128x0.625 mm collimation, a 180 mm FOV, and reconstruction phases at 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90%. We used 50 patients from this dataset. The images have size of 512×512× (205–236) Voxels in DICOM format, and were then converted in NIFTI format. The scans are splitted into training set of 40 scans and validation set of 10 scans.

ICRC participants provided written consent for using their data for research purposes.

The study was conducted in accordance with the Declaration of Helsinki. In UMCU, the study was approved by the local institutional ethics review board (University Medical Center Utrecht, protocol UCC-UNRAVEL #12–387). ICRC participants provided written consent for using their data for research purposes.

ImageCAS dataset

We used 200 scans from the study of Zheng et al. [10]. The scans comprise 3D CTA images from 1000 patients acquired using Siemens 128 slice dual source scanner. For patients with a prior diagnosis of CAD, early revascularisation (within 90 days) is recommended. A high-dose CTA is performed, selecting either the 30–40% phase or the 60–70% phase during the reconstruction to obtain the best possibly coronary artery. The images are of size 512×512× (206 − 275) voxels, with a planar resolution of 0.29–0.43 mm2 and a spacing of 0.25–0.45 mm. The data was collected from real clinical cases at the Guangdong Provincial People's Hospital between April 2012 and December 2018. For more details, please refer to the original paper [10].

The datasets are publicly available, so no approval was necessary for this work.

The regions of interest (ROI_s) that the first and second datasets focus on are the aorta and the coronary arteries. The third one contains just the coronary arteries. The labels include the aorta, right coronary artery, left main coronary artery, left anterior descending coronary artery, left circumflex coronary artery and their vessel bifurcations.

B. Model

As different centers utilize different acquisition protocols, equipment, and parameters, the quality, resolution, and contrast of the resultant images can differ greatly. To ensure that the trained model generalizes well on such data, methods such as data augmentation and normalization must be used to the heterogeneous dataset from multiple centers. These methods help to minimize the variability induced by inter-center differences, allowing the model to perform equally well on data from new centers or devices, which is critical for their practical and clinical use.

Medical equipment can vary in quality and precision, affecting pixel values. Acquisition parameters, including radiation dose, repetition rate, and reconstruction parameters, can also vary [33]. Normalization aids in the correction of these variations and obtaining more uniform, reliable, and interpretable images by mitigating the impacts of variability in equipment and acquisition settings. Normalization entails altering pixel values to ensure compatibility with acquisition parameters. This procedure is required to lessen the impact of variability in equipment and acquisition parameters, allowing for a more reliable interpretation of medical images.

In this work, we used the imaging data without cropping or resizing the images to preserve all the information. The method adopts a 3D UNet architecture, which is easy to use but efficient for segmenting coronary arteries semantically. In particular, the network receives the image and produces the probability semantic map without requiring any complicated processing. The overall structure is a 3D UNet module. The steps involved in the procedure are as follows:

Apply transforms to the images.
Feed the images to the 3D U-Net.
Obtain the prediction results.

On the other hand, the collected coronary masks were subjected to the same preparation in terms of resolution as their corresponding CCTA scans.

Nonetheless, the available memory of the graphics processing unit (GPU) could not accommodate the entire data set and the entire slice at this resolution. Therefore, we divided each scan into sub-volumes with a resolution of 128 × 128 × 128 voxels, saving those that were useful for training purposes by excluding the ones containing non-informative regions outside of the heart area. Subsequently, each sub-volume was sliced into patches with a resolution of 64 × 64 × 64 voxels without overlapping. After segmenting each patch, all patches are combined to reconstruct the segmentation of the entire coronary artery. The collected coronary masks were subjected to the same preparation in terms of resolution as their corresponding CCTA scans.

Figure 1. Example CTTA images from the three datasets used in the study, along with their ground truth segmentations. On the left of each panel, there is an example 2D slice from a 3D CCTA scan; in the middle, the corresponding ground truth segmentation; and on the right, the segmentation overlaid on the CCTA scan. The first two rows correspond to the first dataset, the third and fourth rows to the second dataset, and the last two rows to the third dataset.

Please note that all processing has been performed in 3D. The 2D slices are used for visualization purposes only.

A. Evaluating metric

The segmentation accuracy is compared by calculating the DSC score. The DSC score is calculated as the ratio of twice the area of the correct predictions, i.e., the intersection of predicted and ground truth labels. The ratio is bounded between 0 and 1, with higher values indicating better segmentation accuracy. The formula for DSC is given in (1):

$$\:Dice(G,P\:)=\frac{2*\:Area\:of\:predicted\:correct\:result\:}{\:\text{A}\text{r}\text{e}\text{a}\:\text{o}\text{f}\:\text{p}\text{r}\text{e}\text{d}\text{i}\text{c}\text{t}\text{e}\text{d}\:\text{l}\text{a}\text{b}\text{e}\text{l}\:+\:\text{A}\text{r}\text{e}\text{a}\:\text{o}\text{f}\:\text{g}\text{r}\text{o}\text{u}\text{n}\text{d}\:\text{t}\text{r}\text{u}\text{t}\text{h}\:\text{l}\text{a}\text{b}\text{e}\text{l}\:}\:$$

B. Experimental results

Table 1

Averaged Dice scores of the data comparing the three datasets.
Model 3D U-Net	First Dataset	Second Dataset	Third Dataset
DSC (Train)	0. 952	0.975	0.985
DSC (Test)	0.912	0.939	0.973

The majority of studies are conducted on small databases; expect some papers like [10], and most of them only address the segmentation of the cardiac arteries. In actuality, a radiologist must examine the entire tree containing both the aorta and the arteries in order to make a proper diagnosis of the diseases, such as an aortic aneurysm, aortic stenosis, atherosclerosis, etc., as well as identify anomalies in the origin coronary artery. According to these reasons, we utilized our masks for two distinct purposes. (1) In the first one, the combined mask of the coronary arteries and aorta plays a crucial role in computational fluid dynamics blood flow estimation. By segmenting the aorta and coronary arteries, researchers can obtain continuous 3D measurements of their size and shape. These measurements are essential for objectively evaluating minute changes in aortic morphology over time, allowing for in-depth longitudinal comparisons between various aortic disease states. In the second, cardiologists can particularly benefit from the mask of coronary arteries when assessing the degree of stenosis in areas impacted by CAD. Cardiologists can more accurately determine the degree of stenosis in impacted areas by using the coronary artery mask, which helps with patient care and treatment planning.

The coronary artery and background zone in CCTA images are significantly imbalanced, and the majority of slices provide unnecessary details. Since the coronary arteries only show up as tiny spots in the labels (Figure), we are dealing with sparse data, so we must find a model that can accurately segment the area of interest. However, our model may be confused by the apparent resemblance in appearance between the descending aorta and the surrounding cavities.

When we were training the model with a full image size, we faced many challenges: First, due to the relatively small size and the variation of coronary arteries in the heart from patient to patient, the model struggles to effectively segment the smaller targets. Second, it is not easy to segment the small size of the coronary arteries in the CT scans because they contain a huge amount of information. In addition, the model had difficulty distinguishing between the coronary artery and surrounding area (because the C.A. are small objects) effectively. Last, The GPU cannot accommodate the high-resolution data. For these reasons, we have adopted the patch segmentation approach.

Existing deep learning-based studies and their Dice scores (%) for coronary artery segmentation over the past decade are summarized. Methods can be designed for a wide range of applications, or denotes methods specifically developed and optimized for coronary vessel segmentation.

Table 2

Some Existing works of coronary artery segmentation in the last few years and their Dice scores (in %).
Paper	Year	Methods	Data quantity	Semantic or not	Aorta and /or C A	Modality	Metrics (DSC)
[32]	2017	DeepMedic	42	No	Both	CCTA	75–78
[31]	2018	Two-CNN	44	No	C A	CAG	97
[30]	2018	3D U-Net	18	No	C A	CCTA	95
[29]	2018	Graph Convolutional Network (GCN)	18	No	C A Lumen	CCTA	74
[27]	2019	Four deep learning models	1980 patients	No	C A	CAG	91
[26]	2019	CNN-RNN	150–200 axial slices	No	C A	CCTA	IOU of 84
[25]	2019	Deep learning model based on U-net	3309 patients	No	C A	CAG	91
[24]	2019	3D FCN + level set	70	No	C A	CCTA	90
[23]	2019	3D multi-channel U-Net	33472 training samples	No	C A	CCTA	80
[22]	2020	Global feature embedded network “GFE-Net”	70	No	Both	CCTA	91
[19]	2020	Four deep learning models based on the UNet model	1118 patients	No	CA	CAG	90
[18]	2020	3D FCN	30	No	CA	CCTA	83
[7]	2020	Feature pyramid UNet++	73 patients	Semantic	C A (major vessel)	ICA	66 (LAO) 61 (RAO)
[17]	2021	2D Unet model	69	No	1) Both 2) CA alone	CCTA	91 (for both aorta & CA) 88 (for CA)
[16]	2021	PSPNet network based multi-scale CNN model	109	No	CA	CAG	95
[15]	2021	deep learning and digital image processing	70	No	Both	CCTA	93
[14]	2021	2D and 3D CNNs	59	No	1) Aorta 2) CA	CCTA	97 (For aorta) 86 (For CA)
[13]	2021	3D-Dense-UNet	474	No	CA	CCTA	96
[12]	2022	3D-Unet	243	No	CA	CCTA	77
[11]	2022	MT-Unet++	118	No	CA	ICA	81
[5]	2022	3D-FFR-UNet	68	No	CA	CCTA	82
[10]	2023	Baseline	1000	No	CA	CCTA	82
	2024	Our	70 50 200	Semantic	1) Aorta & CA 2) CA	CCTA	91 94 97

The papers were categorized according to the different types of tasks involved, modalities and the regions of interest (only CA and/or aorta). The first task involves segmentation of the C.A alone in CCTA [31][30][27][25][19][17][15][14][13][12][5][10] and the aorta [23][17][14], in CAG [32][28][26][20][16] and in invasive coronary angiography (fluoroscopy angiography) ICA [18][11]. The second task pertains to semantic segmentation [7], there are some works that have focused on extracting semantic features within a segmentation framework. The networks for coronary artery segmentation are basically based on CNN with its extension like FCN and especially U-net with its variants and extensions, which is specifically designed for semantic segmentation of biomedical images.

In comparison with all works presented in Table2, we can notice that our approach based on patch segmentation provided the best results. It is clear also, that the proposed model performs better in case of multi center and multi vendors devices. We used three different datasets from different countries, and taken in different conditions, protocol acquisition, perspectives, orientations and parameters…etc. we can distinguish that the segmentation method generalize well to any datasets with any conditions according to the training data it was trained on.

In this paper, we have employed a 3D U-NET for coronary arteries semantic segmentation in CCTA data using three datasets with multi center devise. Segmenting the whole volume directly especially a volume with high resolution is poor in term of accuracy. The patching method handle the image of arbitrary size (high or low) which keeping it at its original resolution with preserving crucial features, and making it versatile for various tasks and easy to work on it. Our approach offers better segmentation accuracy in terms of dice coefficient when compared to the segmentation results of other approaches, and the 3D visualization results show how successful the suggested framework is.

Data availability:

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Funding:

P.G., K.L. and V.K. received funding from the European Union’s Horizon Europe research and innovation programme under Grant Agreement No. 101057849 (DataTools4Heart project).”

Contributions

Design of the work: M.T., M.A., S.M. Image collection: P.G., V.K., T.H., K.L. Interpretation of results: M.T, M.A, S.M, K.L. Manuscript draft: M.T., M.A., S.M. Manuscript review: all authors.

Saloni Dattani, Veronika Samborska, Hannah Ritchie and Max Roser (2023) - “Cardiovascular Diseases” Published online at OurWorldInData.org. Retrieved from: 'https://ourworldindata.org/cardiovascular-diseases' [accessed 15 juin, 2024]
Overview of Coronary Artery Disease - Cardiovascular Disorders - MSD Manual Professional Edition (msdmanuals.com)
KOLOSSVÁRY, Márton, SZILVESZTER, Bálint, MERKELY, Béla, et al. Plaque imaging with CT—a comprehensive review on coronary CT angiography based risk assessment. Cardiovascular diagnosis and therapy, 2017, vol. 7, no 5, p. 489.
F. Tian et al. Automatic coronary artery segmentation algorithm based on deep learning and digital image processing. Applied Intelligence, 2021, vol. 51, no 12, p. 8881-8895.
A. Song et al. Automatic Coronary Artery Segmentation of CCTA Images with an Efficient Feature-Fusion-and-Rectification 3D-UNet. IEEE Journal of Biomedical and Health Informatics, 2022.
J. Xin et al. Automatic coronary plaque detection, classification, and stenosis grading using deep learning and radiomics on computed tomography angiography images: a multi-center multi-vendor study. European Radiology, 2022, p. 1-11.
CH. Zhao et al. Semantic segmentation to extract coronary arteries in fluoroscopy angiograms. medRxiv, 2020.
LONG, Jonathan, SHELHAMER, Evan, et DARRELL, Trevor. Fully convolutional networks for semantic segmentation. In : Proceedings of the IEEE conference on computer vision and pattern recognition. 2015. p. 3431-3440.
RONNEBERGER, Olaf, FISCHER, Philipp, et BROX, Thomas. U-net: Convolutional networks for biomedical image segmentation. In : Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer International Publishing, 2015. p. 234-241.
ZENG, An, WU, Chunbiao, LIN, Guisen, et al. Imagecas: A large-scale dataset and benchmark for coronary artery segmentation based on computed tomography angiography images. Computerized Medical Imaging and Graphics, 2023, vol. 109, p. 102287.
H. He et al. Semi-Supervised Coronary Vessels Segmentation from Invasive Coronary Angiography with Connectivity-Preserving Loss Function. In: 2022 IEEE 19th International Symposium on Biomedical Imaging (ISBI). IEEE, 2022. p. 1-5.
Y. LI et al. Automatic coronary artery segmentation and diagnosis of stenosis by deep learning based on computed tomographic coronary angiography. European Radiology, 2022, p. 1-9.
Pan et al. Coronary artery segmentation under class imbalance using a U-Net based architecture on computed tomography angiography images. Scientific Reports, 2021, vol. 11, no 1, p. 1-7.
Gu & Cai. Fusing 2D and 3D convolutional neural networks for the segmentation of aorta and coronary arteries from CT images. Artificial Intelligence in Medicine, 2021, vol. 121, p. 102189.
F. Tian et al. Automatic coronary artery segmentation algorithm based on deep learning and digital image processing. Applied Intelligence, 2021, vol. 51, no 12, p. 8881-8895
X. Zhu et al. Coronary angiography image segmentation based on PSPNet. Computer Methods and Programs in Biomedicine, 2021, vol. 200, p. 105897.
CHEUNG, Wing Keung, BELL, Robert, NAIR, Arjun, et al. A computationally efficient approach to segmentation of the aorta and coronary arteries using deep learning. Ieee Access, 2021, vol. 9, p. 108873-108888
Y. Lei et al. Automated coronary artery segmentation in coronary computed tomography angiography (CCTA) using deep learning neural networks. In: Medical Imaging 2020: Imaging Informatics for Healthcare, Research, and Applications. SPIE, 2020. p. 279-284.
Z. Xian et al. Main coronary vessel segmentation using deep learning in smart medical. Mathematical Problems in Engineering, 2020, vol. 2020.
L. Baskaran et al. Automatic segmentation of multiple cardiovascular structures from cardiac computed tomography angiography images using deep learning. PloS one, 2020, vol. 15, no 5, p. e0232573.
K. Huang et al. Coronary Wall Segmentation in CCTA Scans Via a Hybrid Net with Contours Regularization. In : 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). IEEE, 2020. p. 1743-1747.
J. Gu et al. Segmentation of coronary arteries images using global feature embedded network with active contour loss. Computerized Medical Imaging and Graphics, 2020, vol. 86, p. 101799.
Y. CH. Chen et al. Coronary artery segmentation in cardiac CT angiography using 3D multi-channel U-net. arXiv preprint arXiv:1907.12246, 2019.
Y. Shen et al. Coronary arteries segmentation based on 3D FCN with attention gate and level set function. Ieee Access, 2019, vol. 7, p. 42826-42835.
S. Yang et al. Deep learning segmentation of major vessels in X-ray coronary angiography. Scientific reports, 2019, vol. 9, no 1, p. 1-11.
P. Mirunalini et al. Segmentation of Coronary Arteries from CTA axial slices using Deep Learning techniques. In: TENCON 2019-2019 IEEE Region 10 Conference (TENCON). IEEE, 2019. p. 2074-2080.
S. Yang et al. Major vessel segmentation on x-ray coronary angiography using deep networks with a novel penalty loss function. In: International Conference on Medical Imaging with Deep Learning--Extended Abstract Track. 2019.
J.M. Wolterink et al. Coronary artery centerline extraction in cardiac CT angiography using a CNN-based orientation classifier. Medical image analysis, 2019, vol. 51, p. 46-60.
J.M. Wolterink et al. Graph convolutional networks for coronary artery segmentation in cardiac CT angiography. In : International Workshop on Graph Learning in Medical Imaging. Springer, Cham, 2019. p. 62-69.
W. Huang et al. Coronary artery segmentation by deep learning neural networks on computed tomographic coronary angiographic images. In : 2018 40th Annual international conference of the IEEE engineering in medicine and biology society (EMBC). IEEE, 2018. p. 608-611.
E. Nasr-Esfahani et al. Segmentation of vessels in angiograms using convolutional neural networks. Biomedical Signal Processing and Control, 2018, vol. 40, p. 240-251.
Ø. Kjerland, Segmentation of Coronary Arteries from CT-scans of the heart using Deep Learning. 2017. Thèse de maîtrise. NTNU.
GOLDMAN, Lee W. Principles of CT: radiation dose and image quality. Journal of nuclear medicine technology, 2007, vol. 35, no 4, p. 213-225.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Semantic segmentation of coronary arteries in computed tomography angiograph: A multi-center, multi-vendor and multi disease study

Status:

Version 1

Abstract

Figures

Introduction

Related work

Method

A. Dataset

First dataset

Second dataset

ImageCAS dataset

B. Model

Discussion

CONCLUSION

Declarations

References

Additional Declarations

Status:

Version 1