Automatic Quantitative Coronary Analysis Based on Deep Learning

Abstract

As a core technique to quantitatively assess the stenosis severity of coronary arteries, quantitative coronary analysis (QCA) is urgently supposed to become more automated and intelligent, especially for regions lacking expertise and technology. The existing QCA methods highly depend on manual operation, which is time-consuming and subject to personal experience. This study innovatively proposes a fully automatic QCA workflow based on artificial intelligence (AI-QCA), which can quickly and accurately make a quantitative assessment of stenosis severity. The whole AI-QCA workflow mainly consists of three parts: the boundary-aware segmentation on the coronary angiogram (CAG) images, the AI-enabled coronary artery tree construction, and the diameter fitting and stenosis detection. Experiments show that the precision, recall, and F1 score of the segmentation, evaluated on 1322 CAGs, are 0.866, 0.897, and 0.879, respectively. Furthermore, the RMSE between diameter stenosis assessed by AI-QCA and manual QCA served by senior experts, evaluated on 249 CAGs, is 0.064, and the Pearson coefficient is 0.765. Meanwhile, the operation time can be reduced from tens of minutes to several seconds by AI-QCA. As a conclusion, the proposed AI-QCA is able to quickly quantify stenosis parameters as accurately as senior experts, which is significant for the intelligent diagnosis and treatment of coronary artery disease.

1. Introduction

Coronary artery disease (CAD) is one of the leading causes of mortality in the world. According to the World Health Organization (WHO), in 2021, approximately 17.9 million people died from CAD, accounting for 31% of all global deaths. Furthermore, of these deaths, over 80% occurred in low-income and middle-income countries. Due to the lack of expertise and technology, small hospitals in rural and low-income countries may not be able to properly diagnose and treat CAD patients, which will lead to poor outcomes and high mortality rates. This emphasizes the importance of automatic analysis for coronary imaging data in smart auxiliary medical systems based on artificial intelligence (AI). In this study, we mainly focus on the AI-enabled automatic quantitative coronary analysis (QCA), a crucial technique for the diagnosis of CAD based on coronary angiogram (CAG) images.

In clinical practice, CAG is the gold standard for percutaneous coronary intervention (PCI) surgery. It provides essential diagnostic imaging for interventional cardiologists to understand the patient’s heart condition and make credible treatment decisions. QCA is a set of techniques used to measure the diameter and assess the stenosis degree of the coronary arteries on CAG images, which could promote rational clinical decision-making, risk assessment, and stent placement.

QCA was first introduced in 1977 [1]. With the development of the digital imaging and communications in medicine (DICOM) system and computer image processing algorithms, QCA has become a more and more popular method for quantitatively assessing the results of percutaneous coronary intervention (PCI) surgeries [2,3,4]. Nowadays, the most commonly used QCA systems are CAAS II (PIE Medical, Maastricht, The Netherlands) and QAngio XA (Medis, Leiden, The Netherlands) [5,6,7], both of which have a similar process as follows: First, cardiologists must identify the coronary segment of interest, which focuses on the stenosis or lesion as reference segments. The proximal and distal edges of the coronary area should be relatively free of disease, and the mean value of the diameters of the lumen in these two areas should be calculated as the reference diameter. Then, based on the location of the area of interest, analysts manually trace the centerline, and edge detection algorithms are used to identify the margins of coronary artery segments. A series of diameters are calculated along the entire segment, and a trend line of the luminal diameters is drawn. Additionally, an interpolation algorithm is applied to the region between the generated margin line and a hypothetically normal coronary margin, highlighting the area of stenosis [8]. As a result, several parameters [9] are figured out by the QCA system: (1) lesion length (LL, mm), defined as the length of a coronary segment of interest between the proximal point and distal point; (2) minimal luminal diameter (MLD, mm), defined as the smallest diameter of lumen; (3) reference vessel diameter (RVD, mm), defined as the average diameter of coronary lumen free from the atherosclerotic illness; and (4) diameter stenosis (DS, n%), defined as (RVD- MLD)/RVD.

However, the QCA process and the existing systems heavily rely on manual operation, which is time-consuming and can even take cardiologists up to 20–30 min to complete accurate analyses for only one single CAG. Although many researchers have tried to improve corresponding algorithms by using various filter-based methods, such as the Hessian matrix-based features [10], radon-like features [11], and Gabor wavelet features [12], these methods are not efficient for clinical applications due to their complex computation and pixel-wise operations. Fortunately, with the advancements in artificial intelligence (AI) and computer vision in the field of medical imaging, QCA algorithms have the potential to be more automated and intelligent.

In recent years, there has been a growing interest in using AI techniques to analyze CAG. Cong et al. [13] proposed a deep learning-based workflow for stenosis detection and severity classification, and Danilov et al. [14] combined three models with various architectures to raise the accuracy. Ovalle [15] applied quantum computing in the context of a hybrid transfer-learning paradigm for stenosis detection to improve the performance of a pretrained network. Additionally, Tmenova [16] applied CycleGAN to simulate angiograms for the purpose of augmenting datasets. Furthermore, a novel approach known as dynamic coronary road mapping with deep learning-based Bayesian filtering has been reported to improve visual feedback and reduce the use of contrast during PCI [17]. Papandrianos [18] used an RGB-CNN model for SPECT myocardial perfusion imaging to diagnose CAD and compared it with pretrained VGG-16 and DenseNet-121 networks. Wang [19] investigated a deep-learning algorithm for the quantification of coronary artery calcium scores based on computed tomography data from 530 patients. Actually, the most common application of deep learning in CAG is coronary artery segmentation. Esfahani [20] proposed a convolutional neural network (CNN) approach for detecting vessel regions in angiography images. With the development of a fully convolutional network (FCN) and U-Net, deep learning shows good performance in coronary artery segmentation. Yang et al. [21] proposed a robust method for major vessel segmentation that could maintain high connectivity in most narrow areas. Later, they introduced a penalty term for penalizing false negatives and false positives into the dice coefficient of the loss function to improve performance [22]. Baskaran [23] evaluated an end-to-end U-Net-inspired deep learning model for the segmentation and quantification of cardiac structures in coronary computed tomography angiography. Andrushia [24] presented a framework for a new segmentation model for leukocyte images using an extreme learning machine, which can assist in diagnosing various diseases, such as leukemia, malaria, psoriasis, and AIDS. In order to further stimulate the potential of deep learning in segmentation, more and more research has combined different machine learning techniques into deep learning. Gao [25] applied the gradient boosting decision tree (GBDT) and deep forest classifiers into CAG segmentation, while Mulay [26] proposed an adaptive instance normalization style transfer technique for segmenting the coronary arteries.

However, most works focus on improving the overall accuracy of segmentation instead of emphasizing the edge information of the coronary artery, which just ignores one of the most crucial factors for accurate QCA computation and thus cannot be competent in the process of QCA. Additionally, previous works [21,22,27], in this field, primarily focused on the segmentation of main vessels, including the left anterior descending (LAD), left circumflex artery (LCX), and right coronary artery (RCA), but neglected their primary branches. These branches, defined as vessels greater than 1 mm, are also clinically significant and important for QCA. Differently, we focus on the segmentation and quantification of not only the primary coronary arteries, including LAD, LCX, and RCA, but also their primary branches, where edge information about coronary arteries is also well considered.

Overall, all the above literature have made significant contributions, but the studies on automatic QCA are very few. Only several studies have attempted to incorporate deep learning techniques into QCA to improve its accuracy and efficiency. Busto [28] proposed an automatic QCA method using a U-Net segmentation model. However, the QCA workflow was not entirely automated, as the stenosis lesion areas were manually selected and the segmentation results required manual correction. Zhao [29] integrated a feature pyramid with a U-Net++ model for automatic extraction and stenosis evaluation. However, the dataset used for segmentation was limited, resulting in many redundant branches, and the results merely referred to the classification of stenosis degrees instead of quantification. Hong [30] used deep learning to quantify coronary artery disease according to computed tomography angiography images from 156 patients. Similar to the work in [28], Hong’s method was not entirely automated, as the majority was realized by a semi-automated software named Autoplaque, where manual operations were required.

In this study, we propose a fully automatic workflow for a QCA-based deep learning framework, named AI-QCA, with the purposes of achieving an accurate assessment of CAD in an intelligent manner, minimizing the time cost, and eliminating manual intervention. Our proposed QCA workflow mainly consists of three parts: First, a boundary-aware segmentation architecture is proposed to segment the main vessels and primary branch vessels. Then, the coronary artery tree is constructed on the basis of deep learning models, where the root node is auto-located via a proposed searching algorithm. Finally, by applying techniques such as Gaussian smoothing, branch separation, reference diameter fitting, and stenosis detection, the system can automatically generate a fitting line for the diameters and detect stenosis regions with quantitative parameters. The main advantages of our proposed AI-QCA can be summarized in three aspects as follows: (1) We realize a fully automatic QCA process, which can effectively avoid the interference of personal preference and experience. (2) AI-QCA is able to quantify stenosis parameters as accurately as the senior experts, benefiting from all of the boundary-aware segmentation based on AI, the proposed methods of coronary artery tree (CAT) construction, and the efficient utilization of fitting algorithms. (3) AI-QCA can reduce the time cost from tens of minutes to several seconds compared with manual QCA. All the above is significant for the intelligent diagnosis and treatment of coronary artery disease.

2. Materials and Methods

2.1. Overview of AI-QCA

The proposed AI-QCA mainly consists of three stages. Section 2.3 introduces the BasNet [31] for binary vessel segmentation with more attention to edge information. Then, a root node location workflow based on PSPNet [32] is presented in Section 2.4, and the CAT construction is illustrated on the basis of Zhang’s parallel thinning algorithm [33] and root node searching algorithm. In Section 2.5, the AI-QCA algorithms are applied to generate quantitative parameters and fitting results for QCA, including Gaussian smoothing [34], branch separation, reference diameter fitting, and stenosis detection. The workflow of the whole process is shown in Figure 1.

2.2. Study Population and Image Annotation

Overall, 13,222 angiograms were retrospectively collected from 3275 patients who underwent X-ray coronary angiography performed at Fuwai Hospital, 11,900 of which were used for training a vessel segmentation model, and 1322 angiograms were used in the validation. Additionally, another 249 angiograms were used to evaluate the performance of AI-QCA, including 139 LAD, 36 LCX, and 74 RCA. Figure 2 shows the study flowchart. The vessel segmentation annotation was on a pixel-by-pixel level. All coronary arteries with a diameter of at least 1 mm were visually identified and marked by cardiological analysts on the original angiograms, and then a team of trained and certified technicians meticulously labeled each pixel based on the initial markings. The study was approved by the Institutional Review Board of Fuwai Hospital. All the angiograms were processed by information desensitization algorithms.

2.3. Boundary-Aware Coronary Artery Segmentation

The QCA system demands precise measurements of the vascular centerline and vessel diameter, both of which are closely tied to the boundary information of the segmentation results. In the past, segmentation models primarily focused on achieving high accuracy of the overall regions being segmented, but in order to obtain more accurate diameter information, BasNet was implemented as the coronary artery segmentation model. The whole architecture of BasNet includes a prediction module and a refinement module, representing a process from coarsening to refining with the purpose of capturing both high-level global information and low-level detailed information (Figure 3). The predict module is an encoder-decoder network similar to U-Net [35] that learns predictive feature maps from input images. In order to reduce overfitting, each layer of the decoder is supervised by ground truth. To further capture global information, bridge stages are added between the encoder and decoder. The refinement module is a simpler encoder-decoder module that refines coarse maps obtained by the prediction module. The refinement process learns the differences between coarse maps and the ground truth by analyzing residual blocks. Eight side outputs were supervised in the architecture, including seven side outputs in the prediction module and one side output in the refine module. The overall loss is the weighted sum of the eight losses.

$$\mathrm{L}={{\displaystyle \sum }}_{\mathrm{k}=1}^{\mathrm{K}}{\mathsf{\alpha }}_{\mathrm{k}}{\mathrm{l}}^{\left(\mathrm{k}\right)}$$

(1)

For each loss function of every side output, it contains three types of different losses. On the pixel level, the binary cross entropy loss (BCE Loss) [36] is conducive to the convergence of all pixels, ignoring the label of surrounding pixels, which is the most common loss function in segmentation tasks. The BCE Loss function is shown as follows, where G(r,c)∈(0,1) represents the true value in row r and column c, and S(r,c) represents the predicted value.

$${\mathrm{L}}_{\mathrm{bce}}=-{{\displaystyle \sum }}_{(\mathrm{r},\mathrm{c})}[\mathrm{G}\left(\mathrm{r},\mathrm{c}\right)\log \left(\mathrm{S}\left(\mathrm{r},\mathrm{c}\right)\right)+(1 - \mathrm{G}\left(\mathrm{r},\mathrm{c}\right))\log \left(1 - \mathrm{S}\left(\mathrm{r},\mathrm{c}\right)\right)]$$

(2)

On the patch level, the structural similarity (SSIM) loss [37] could enhance the relationship of each pixel and the local patch, which helps to strengthen the loss value in edge parts of the object and suppress nonedge information. Therefore, the algorithm can pay more attention to the marginal details of the interested vessels. The SSIM loss function is shown as follows, where μ_x, μ_y and σ_x, σ_y represent the mean and variance of x and y, σ_xy is the covariance of x and y, while C₁ and C₂ are constants to avoid the case of a denominator equal to 0. In general, C₁ = (K₁/L)², C₂ = (K₂/L)², where K₁ = 0.01, K₁ = 0.03, and L = 255 is the maximum pixel value in the image.

$${\mathrm{L}}_{\mathrm{ssim}}=1 -\frac{{(2\mathsf{\mu }}_{\mathrm{x}}{\mathsf{\mu }}_{\mathrm{y}}{ + \mathrm{C}}_1{\left)\right(2\mathsf{\sigma }}_{\mathrm{xy}}{ + \mathrm{C}}_2)}{{(\mathsf{\mu }}_{\mathrm{x}}^2{ + \mathsf{\mu }}_{\mathrm{y}}^2{ + \mathrm{C}}_1{\left)\right(\mathsf{\sigma }}_{\mathrm{x}}^2{ + \mathsf{\sigma }}_{\mathrm{y}}^2{ + \mathrm{C}}_2)}$$

(3)

On map level, intersection over union loss (IoU Loss) [38] is used to measure the similarity of two sets, which gets the foreground more attention. The following is IoU loss, where S(r,c) and G(r,c) represent the predicted value and true value in row r and column c, respectively.

$${\mathrm{L}}_{\mathrm{iou}}=1 - \frac{{\sum }_{\mathrm{r}=1}^{\mathrm{H}}{\sum }_{\mathrm{c}=1}^{\mathrm{W}}\mathrm{S}\left(\mathrm{r},\mathrm{c}\right)\mathrm{G}\left(\mathrm{r},\mathrm{c}\right)}{{\sum }_{\mathrm{r}=1}^{\mathrm{H}}{\sum }_{\mathrm{c}=1}^{\mathrm{W}}[\mathrm{S}\left(\mathrm{r},\mathrm{c}\right)+\mathrm{G}\left(\mathrm{r},\mathrm{c}\right) - \mathrm{S}\left(\mathrm{r},\mathrm{c}\right)\mathrm{G}\left(\mathrm{r},\mathrm{c}\right)]}$$

(4)

The overall loss comprehensively considers different levels of the significance information, which can not only maintain high accuracy for the segmentation results, but also learn the structure of the original blood vessel by strengthening the boundary loss.

2.4. Root Node Automatic Location Algorithm and CAT Construction

In the QCA system, the centerline is the basis for acquiring vascular structure and calculating vascular diameter. In our work, Zhang’s fast parallel thinning algorithm is adopted to extract the centerline iteratively according to the binary segmentation result. Until a skeleton with a width of one single pixel is obtained, the algorithm continuously removes pixels from the outer layer of vessels but never affects the overall shape. However, the simple centerline extraction algorithm does not contain information on vascular diameter and direction, which is the main reason why many studies require manual participation. In order to solve this problem and avoid manual participation, this paper presents a set of root node automatic location algorithms. Firstly, a deep learning network is designed to automatically find the root node. It is a three-value segmentation network based on PSPNet, which is used to classify the areas of the catheter, the blood vessel, and the starting point area of the blood vessel (Figure 4). The annotations used in this study are based on our prior work [39], which aimed at multiclass vessel segmentation and detailed classification of coronary vessels. In this study, only the catheter and the starting part of the coronary artery segmentation were annotated separately, while other classes of coronary arteries were grouped into the same annotated category. This algorithm provides valuable support and facilitates great improvements for the subsequent process. On the one hand, segmenting the catheter in the images is crucial for accurate scale conversion from pixels to millimeters, as the size of the catheter remains constant in millimeters despite variations in pixels across different CAGs. In this study, the diameter of the catheter used for coronary angiography was fixed at 1.67 mm. On the other hand, the segmentation results help to confirm the starting point for the main vessel segmentation and centerline extraction by using a searching algorithm for root nodes.

Based on the results of the root node-assisted network, the root node can be searched by Algorithm A1. The specific steps of the root node searching algorithm are shown in Appendix A.

Once the root node and the pixels-to-millimeters scales were determined, the CAT could be built iteratively along the centerline. Every node of the CAT stores information, including its coordinates, the coordinates of the next node, the direction of the vessel, and the diameter of the vessel at that node. In order to calculate the diameter, considering the current node as the center, a circle with a gradually increasing radius is drawn, making sure that the entire circle is in the segmented vessel. When the circle is exactly tangent to the edge of the vessel, the diameter of the circle is considered the diameter of the vessel at this node.

2.5. AI-QCA

On the basis of the segmentation results and generated CATs, a series of fitting and optimization algorithms were designed to automatically detect and quantify the stenosis lesions by making full use of the diameter information of the CAT. The whole process generates a diameter fitting diagram, marks the stenosis areas on the original CAG, and calculates the LL, MLD, RVD, and DS in every detected stenosis area.

2.5.1. Gaussian Smoothing

The Gaussian smoothing technique is applied to reduce noise and smooth the data in the series of diameter information from the CAT. The Gaussian kernel, controlled by a mean value and standard deviation, is convolved with the image by sliding over it and computing a weighted average of pixel values within the kernel’s footprint. In this work, with the comparative experiments of different sets of parameters, the kernel size is set to 21, and the standard deviation is set to 1.5. The optimization of the algorithm, due to the symmetry of the vessel centerline, reduces the calculation to only half of it.

2.5.2. Branch Separation

In the processing of segmentation and QCA, the concerned vessels were always the coronary arteries, including LAD, LCX, RCA, and their primary branches. The resulting CAT data often includes multiple branches. However, when analyzing the stenosis region in the process of QCA, it is important to focus on a single vessel pathway. Therefore, a branch separation algorithm is necessary. A simple method is presented to distinguish left and right branches below the bifurcation point. For three points on the CAT, assumed as P₁ = (x₁, y₁), P₂ = (x₂, y₂), and P₃ = (x₃, y₃), P₁ is a point before the bifurcation point, P₂ is the bifurcation point, and P₃ is a point after the bifurcation point. The amount of area of three points in the plane is calculated as follows:

$$\mathrm{S}=(x_1 - x_3) \times (y_2 - y_3) - (y_1 - y_3) \times (x_2 - x_3)$$

(5)

P₁, P₂, and P₃ are clockwise when S is positive, counterclockwise when S is negative, and collinear when S is 0. According to the results of direction judgment, the branches below the bifurcation point can be distinguished as left or right. By combining the diameter value trends and segmentation results, we can identify the branches that are closest to the extension direction and have diameter values of the main vessels before bifurcation. This allows us to separate the main branches and their primary branches.

2.5.3. Reference Diameter Fitting

In order to accurately determine the reference diameter at each location, a linear regression method is employed to fit the reference diameter line. Unlike traditional QCA methods that use only proximal and distal diameters to calculate the reference diameter line, this approach takes advantage of all data points of vessel diameters, thereby reducing potential errors. The least squares algorithm [40] is employed to find the line that best fits the data by minimizing the residual sum of squares (RSS) between the reference diameters and the actual diameters.

$$\mathrm{RSS}={\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{m}}{{(\mathrm{y}}_{\mathrm{i}} {- \hat{\mathrm{y}}}_{\mathrm{i}})}^2$$

(6)

However, there were some exceptions that needed to be excluded for global fitting. To optimize the fitting method, some additional conditions need to be taken into account. For example, when there is a cliff-like change in the diameter of a segment of the blood vessel, it should be excluded from the global fitting process. Additionally, the data used for global fitting are filtered by threshold values. For example, a threshold of 0.3 is set to participate in global fitting because segments of vessels with a stenosis degree greater than 0.3 are considered potential stenosis areas and should not be used as reference diameters. Additionally, the starting and ending areas of the vessel are excluded due to inaccuracies. The diffusion of the contrast agent from the catheter will mislead the visible diameter in the starting area, while the errors will be larger for the low diameter values in the ending area. Additionally, bifurcation points are also excluded. Overall, Figure 5 shows the separated branches of the CAT and generated fitting reference diameters in the CAG in red lines, while the different areas are filled in yellow, indicating that stenosis lesions may exist.

2.5.4. Stenosis Detection

According to the reference diameters, the rate of stenosis at each position on the CAT can be calculated. To detect stenosis, we first removed data with reference diameters less than 1 mm, because these vessels are typically located at the end of the vessel and are of relatively little medical significance for PCI. Then, the vessels are segmented by specific points, including the maximum, minimum, bifurcation, and adjacent points for reference diameters. This allows us to extract multiple effective segments and detect the narrowest position of each segment.

The segments are then merged to accurately determine the proximal and distal areas of the stenosis. A stenosis uniqueness validation is performed to ensure that the current position is the narrowest among the potential stenosis areas nearby. It is not sufficient to only consider the narrowest points at similar candidate segments, but it is also important to evaluate the distance from existing stenosis to avoid missing any detection positions. Additionally, the DBSCAN [41] algorithm is used to eliminate duplicated coordinates of detection results by clustering within small segments. The potential stenosis candidate segments may have overlapping areas with adjacent segments, resulting in close proximity of the coordinates in these areas. However, the coordinates may not be exactly the same. To address this issue, DBSCAN can be utilized to eliminate the redundant coordinates and retain the validity of the remaining coordinates. After the stenosis coordinates are determined, the segments that contain the stenosis are merged, and the DS, MLD, RVD, and LL at the coordinates are calculated as the output of the AI-QCA.

3. Results

3.1. Segmentation Evaluation

3.1.1. Experimental Setup and Evaluation Metrics

In order to access the performance of segmentation on LAD, LCX, and RCA with primary branches, 1322 is involved in the evaluation. In the training process, the Adam optimizer is adopted with an initial learning rate of 0.001, and the batch size is set to eight. These models were implemented using pyTorch 1.10.1, and they were trained for 20 epochs and tested on an NVIDIA TITAN RTX 24GB GPU.

The metrics for segmentation performance include accuracy, precision, recall, F1 score, and Jaccard coefficient. The definitions of the metrics are as follows: accuracy = (TP + TN)/(TP + TN + FP + FN); precision = TP/(TP + FP); recall = TP/(TP + FN); F1 score = 2 × TP/(2 × TP + FP + FN), which is equal to the dice coefficient; Jaccard coefficient = TP/(TP + FP + FN), which means the intersection over union (IoU) of the predicted vessels. TP is true positive, FP is false positive, TN is true negative, and FN is false negative.

3.1.2. Performance of Segmentation

In this work, several segmentation models were adopted to assess the segmentation performance for comparison experiments, including BasNet, U²-Net [42], Pix2Pix [43], and simple U-Net.

Table 1. Segmentation performance for all vessels.

Method	Accuracy	Precision	Recall	F1 Score	Jaccard
U-Net	0.984	0.777	0.818	0.788	0.663
U2-Net	0.985	0.852	0.731	0.772	0.653
Pix2Pix	0.988	0.798	0.875	0.828	0.720
BasNet	0.991	0.866	0.897	0.879	0.787

,

Table 2. Segmentation performance for LAD with primary branches.

Method	Accuracy	Precision	Recall	F1 Score	Jaccard
U-Net	0.985	0.781	0.816	0.791	0.663
U2-Net	0.989	0.863	0.820	0.836	0.723
Pix2Pix	0.989	0.796	0.921	0.851	0.745
BasNet	0.991	0.865	0.896	0.878	0.786

,

Table 3. Segmentation performance for LCX with primary branches.

Method	Accuracy	Precision	Recall	F1 Score	Jaccard
U-Net	0.978	0.669	0.773	0.703	0.557
U2-Net	0.987	0.861	0.742	0.789	0.663
Pix2Pix	0.983	0.754	0.755	0.741	0.618
BasNet	0.991	0.861	0.893	0.874	0.780

and

Table 4. Segmentation performance for RCA with primary branches.

Method	Accuracy	Precision	Recall	F1 Score	Jaccard
U-Net	0.989	0.861	0.858	0.855	0.751
U2-Net	0.980	0.836	0.642	0.701	0.583
Pix2Pix	0.990	0.834	0.930	0.877	0.784
BasNet	0.991	0.872	0.900	0.884	0.794

show the metrics for different types of coronary arteries, specifically referring to all vessels, LAD, LCX, and RCA, respectively. As shown in Table 1, Table 2, Table 3 and Table 4, BasNet statistically outperforms other models in most of the metrics for every type of vessel, except Pix2Pix has better performance in recall in terms of LAD and RCA. The F1 score of BasNet for all vessels can reach up to 89.7%.

Several representative examples of different segmentation models on different types of vessels are displayed in Figure 6. According to the visual comparison, the results of the BasNet are closer to ground truth, while there are more redundant or lacking branches in other models. In addition, BasNet exhibited better connectivity on the bifurcation position and stenosis area.

3.2. Performance of CAT

CAT construction is a crucial step of AI-QCA, which is closely associated with the performance of segmentation. In order to evaluate the performance of CAT construction, the average distance is proposed to measure the difference between the generated CAT and the ground truth of CAT. In some sense, the performance of CAT construction also represents the edge effect of segmentation, because CAT derives from the centerline of the vessel boundary sides. Due to the existence of different lengths between the two CATs, the dynamic programming (DP) algorithm [44] was adopted to align the vessels by point-to-point matching (Figure 7).

According to the vessel match by the DP algorithm, the average distance variances of corresponding vessels can be calculated. The average distance variances of CAT generated by BasNet and Pix2Pix are listed in

Table 5. The average distance variances of CAT.

Vessels	BasNet (mm)	Pix2Pix (mm)
LAD main	0.515	0.565
LAD branch	0.962	1.089
LCX main	0.548	0.600
LCX branch	1.176	1.307
RCA main	0.302	0.304
RCA branch	1.303	1.418
overall	0.944	1.011

CAT = coronary artery tree; LAD = left anterior descending; LCX = left circumflex artery; RCA = right coronary artery.

. The overall distance variances of CAT generated by BasNet and Pix2Pix were 0.944 and 1.011, respectively. For every type of vessel, the distance variances generated by BasNet are lower than those of Pix2Pix, which means the CAT construction of BasNet behaves better than Pix2Pix. The average distance variance for the RCA main was the lowest among all types of vessels, illustrating that the RCA main is more likely to have accurate CAT results. Figure 8 shows the visualization of the distance variance along the CAT from the root node. It can also be seen that the distances measured by BasNet are mostly lower than those measured by Pix2Pix.

3.3. Evaluation of AI-QCA

DS, MLD, RVD, and LL were evaluated on AI-QCA compared to QCA. The QCA results are generated from the reports by experts using the Medis QAngio XA 7.3 system. Figure 9 shows the comparison of AI-QCA and QCA, with root mean square error (RMSE) and Pearson coefficient taken as the measurements. As a result, the RMSE of DS, MLD, RVD, and LL are 0.064, 0.306, 0.421, and 3.984 for all vessels, 0.063, 0.321, 0.480, and 4.299 for LAD, 0.061, 0.290, 0.370, and 3.010 for LCX, 0.058, 0.281, 0.312, and 3.650 for RCA, respectively. The Pearson coefficients of DS, MLD, RVD, and LL are 0.765, 0.700, 0.664, and 0.777 for all vessels, 0.771, 0.742, 0.667, and 0.776 for LAD, 0.821, 0.725, 0.667, and 0.785 for LCX, and 0.802, 0.647, 0.782, and 0.763 for RCA, respectively. All the p-values are lower than 0.001.

Examples of AI-QCA applied to LAD, LCX, and RCA are shown in Figure 10, Figure 11 and Figure 12. According to the comparison of (a) and (b) in the figures, the stenosis areas, and fitting lines of AI-QCA are comparable to those of QCA, with similar fitting CAG and fitting diameter. The DS, MLD, RVD, and LL are also very close. The comparison of (a) and (b) in the figures demonstrates that the stenosis areas and fitting lines generated by the AI-QCA are consistent with those produced by the QCA, exhibiting similar fitting CAG and fitting diameters. Additionally, the DS, MLD, RVD, and LL values obtained from the AI-QCA are also in close agreement with those generated by the QCA. Furthermore, the comparison of (c) and (d) shows that the diameter curves and diameter fitting lines produced by AI-QCA are similar to those obtained by manual QCA, and the locations of stenosis areas generated by the AI-QCA (proximal, ostial, and distal) are also highly consistent with those generated by experts. Due to its fully automated nature, the AI-QCA process reduces labor costs and saves time, as it can analyze a CAG in just 8 s, while the manual QCA process can take up to 20–30 min to complete.

4. Discussion

In this study, we present a new automated stenosis analysis method called AI-QCA, which utilizes deep segmentation models. The proposed AI-QCA achieved several goals, as follows: Firstly, AI-QCA achieves fully automatic stenosis analysis and generates quantitative parameters, including DS, MLD, RVD, and LL, without any manual operation in the process. Secondly, according to the low values of RMSE, the generated quantitative parameters are in high agreement with the QCA results operated by experts. Thirdly, the AI-QCA significantly reduces the time cost of stenosis analysis, from 20 min with a manual QCA to just 8 s with AI-QCA.

For segmentation, Figure 6 shows the intuitive visual results of BasNet, Pix2Pix, U²-Net, and U-Net. BasNet and U²-Net have better performance in connectivity, while U-Net sometimes leads to the disconnection of vessel branches. More branch lacking occurred in U²-Net, and Pix2Pix may cause more small redundant branches. Overall, BasNet shows the best segmentation performance, which is closer to ground truth.

The results of Table 1, Table 2, Table 3 and Table 4 interpret the statements above. Except for BasNet, Pix2Pix behaves better in recall, while U²-Net behaves better in precision. For vessel types, Pix2Pix behaves better in terms of LAD and RCA than U²-Net, while it is just the opposite for LCX. The underlying reason is that Pix2Pix is an image-to-image translation method based on generative adversarial networks (GAN), which tends to segment more vessels, contributing to higher recall, but it also comes with more small redundant branches. U²-Net is a deep segmentation approach with nested U-structure using inverted residual block and dual-channel attention mechanism, which concentrates more on improving the precision, causing the occasional loss of vessel branches. While BasNet takes advantage of multiscale feature fusion mechanisms and handles different levels of saliency, it has better segmentation performance.

In the evaluation of CAT, the average distance variances from BasNet are obviously lower than those of Pix2Pix, according to Table 5. This is because BASNet is specifically designed to be boundary-aware, whereas the points on the CAT represent the average of the vessel edges. It can be deduced from Figure 8 that distance variances increase along the CAT. The distance variances of the main vessels are relatively low, but tend to increase when it comes to the branches, regardless of the type of vessel. The distance variances of the main vessels are similar for BasNet and Pix2Pix because the segmentation performances are both excellent. However, BasNet outperforms Pix2Pix in the segmentation of branches, as seen in the lower distance variances when the points on the CAT are far from the root node.

In the QCA evaluation, DS, MLD, RVD, and LL generated by AI-QCA showed agreement with those of QCA operated by experts, as shown by the RMSE values in Figure 9. At the same time, we can see that the Pearson coefficients of DS are better than MLD, RVD, and LL. Additionally, a small amount of MLD, RVD, and LL in AI-QCA are higher than those in QCA. The first reason may be attributed to the errors with the scales caused by converting pixels to millimeters as defined by the segmentation of catheter width. In general, the catheter width only occupied 6–10 pixels, which means even a one-pixel error in the results of segmentation will lead to large errors in millimeters. In fact, the segmentation of particularly fine objects inherently has some deviation, especially at the edge positions where pixel values are relatively low, which can easily be neglected by the model, and result in an overestimated scale. DS is a percentage that is unrelated to the scales, while the scale conversion of MLD, RVD, and LL may amplify the errors. Therefore, a larger reference for scale calculation is supposed to be taken into account to reduce the error in the future. Additionally, the value of LL derives from the merging of small stenotic regions, which requires improving the rationality of the parameters in interval merging and boundary setting. Additionally, the LL of manual QCA may be smaller because people may choose a boundary position of the stenosis interval with very small stenosis lesions by mistake, similar to the proximal location in Figure 12c.

In fact, there are also limitations to AI-QCA. Firstly, the current algorithm is heavily dependent on the quality of the image and the performance of the segmentation model. In order to improve its accuracy, it is important to include more high-quality training images in the training process. Secondly, only one frame of CAG is involved, while incorporating dynamic frames from DICOM can provide more information for better image analysis. Furthermore, a single view of CAG may cause incorrect estimation of diameters if the stenosis lesion is an eccentric stenosis. Incorporating results from different views may provide more detailed and accurate information in 3D space for stenosis lesion analysis.

In the future, AI-QCA has the potential to be used to help cardiologists rapidly locate the stenosis area and acquire quantitative results for reference. To achieve this, further assessment in external clinical contexts is necessary to test its practicality. In addition, AI-QCA has the potential to be integrated into the SYNTAX scoring system [45], a commonly used tool to evaluate the complexity of CAD and determine the optimal course of treatment. By incorporating AI technology into the SYNTAX score, the accuracy and efficiency of this system could be significantly improved.

5. Conclusions

In this study, a fully automatic AI-QCA system based on deep learning is proposed. This AI-QCA system produces quantitative stenosis parameters with high accuracy in a matter of seconds, significantly reducing the time and labor costs for cardiologists. The experiment results on a clinical dataset show that the AI-QCA has strong consistency with traditional QCA methods.

It is noted that the results of this study may vary from previous results for several reasons. Firstly, the specific deep learning algorithms and techniques used in the proposed AI-QCA system may differ from those used in previous studies, which focused more on boundary information. Additionally, the segmentation annotation data involving the primary branches may be different from that used in previous studies. Furthermore, the evaluation metrics and methods used to assess the performance of the AI-QCA system may also vary between studies, which can also contribute to differences in results.

In conclusion, the proposed AI-QCA system demonstrates promising potential as a fast and accurate method for quantifying stenosis parameters in coronary angiography images. In the future, further evaluations in larger and more diverse patient populations will be necessary to fully validate the performance and clinical utility of this system.