Datasets
Publicly available dataset is used in this study that collected from the Mendeley website. In this dataset, it was present a set of angiographic imaging series who underwent coronary angiography using Coroscopy (Siemens) and Innova (GE Healthcare) image-guided surgery systems at the Research Institute for Complex Problems of Cardiovascular Diseases (Kemerovo, Russia). All patients had angiographically and/or functionally confirmed one-vessel coronary artery disease (≥ 70% diameter stenosis by quantitative coronary analysis or 50–69% with FFR (fractional flow reserve) ≤ 0.80 or stress echocardiography evidence of regional ischemia). Angiographic images of the radiopaque overlaid coronary arteries with stenotic segments were selected and then converted into separate images. An interventional cardiologist rejected non-informative images and selected only those containing contrast passages through a stenotic vessel. A total of 1934 grayscale images (20 patients) of 512×512 to 1000×1000 pixels were included in the dataset [32]. All collected image are annotated from the imglab tool. This tool is very useful for labelling the image for the yolov8 model. After the labelling the dataset, it is splited into three part such as training, testing and validation according to yoloV8 model.
Pre-processing
The angiography images of heart vessels were pre-processes the images using the set of various image processing techniques. The pre-processing steps for coronary angiography images play an important role in preparing the dataset for analysis and diagnosis of blockage of vessels. Augmentation helps in creating a diverse dataset by applying various transformations model to the existing images. In the context of coronary angiography, augmentation techniques may include rotation, scaling, flipping, and translation. And also handle the variations in patient positioning and image acquisition. Contrast enhancement improves visibility of structures within the image by adjusting the intensity levels. In coronary angiography, enhancing the contrast is crucial for highlighting blood vessels and identifying potential blockages or abnormalities. Techniques such as histogram equalization, contrast stretching, and adaptive histogram equalization can be applied to enhance the visibility of fine details in angiography images. Noise reduction ensures image quality by suppressing unwanted noise while preserving essential details. Filters such as Gaussian filters, median filters, and wavelet de-noising are commonly employed to decrease the noise in coronary angiography images. And last part is normalization means scaling the pixel values of a standard range, making them consistent across different images. This step is essential for ensuring that the input data has a uniform distribution, which is particularly important for machine learning algorithms. Normalization can help in improving the convergence of training algorithms and ensures that the models are not biased towards specific intensity ranges present in the original images. These proposed model work on the deep learning YOLOv8 architecture. Firstly, we are applying the data pre-processing, data augmentation of the dataset and using the deep learning models to evaluate the detection of blockage of heart vessels. The steps for the blockage detection of heart vessels have been given in Fig. 1.
YoloV8 Model
The proposed framework for coronary vessels blockage detection is based on the YOLOv8 architecture. YOLOv8 model is an improvement of the YOLOv5 model. This model uses a large number of backbones and used feature fusion methods for achieve more precise accuracy in term of object detection techniques. The model of YOLOv8 are described using this blood vessels of angiography image as follows
From Fig. 1, YOLOv8 mostly uses images will be fixed size during training and inference. Commonly required sizes of images include 416x416, 608x608, and 1024x1024 pixels respectively. The Smaller sizes resultant images in faster inference but may sacrifice detection accuracy, while larger sizes of resultant images improved detection accuracy. YOLOv8 architecture is utilize multi-scale feature fusion techniques to integrate features from different layers of the backbone network, allowing the model to capture objects of various sizes and scales more effectively. In above model, both point wise and depth wise convolutional layers are playing important task for extracting the feature the input images and provide the facility to detect the object. Point-wise convolutions are worked in YOLOv8 architecture after feature maps with higher spatial resolution. It helps to decrease the computational cost by reducing the number of channels. After the complete the processes of Point wise convolutions, it followed to depth wise convolutions or other convolutional operations to further process the feature. Depth wise convolutions are primarily used in YOLOv8 model for efficient feature extraction to achieve the accurate blockage detection of heart vessels.
Evaluation
In this paper, we have to evaluate the confusion metrics for evaluate the performance of deep learning (YOLOv8 architecture) model. The performance of proposed learning model of coronary artery blockage detection evaluated using various confusion metrics, including with precision, recall, and mAP (Mean Average Precision) score form Eqs. 1,2 and 3 which is described in below. The results of the evaluation were compared to those of other deep learning models reported in the literature.
Precision
It is a measurement of a model's predictions with respect to positive instances. It quantifies the percentage of actual true positive predictions out of total predicted positive and false positive instances. Precision is commonly used as an evaluation metric in binary classification tasks [33]. A larger precision value identifies some false positives, which means the model is predicting some incorrect positive predictions. Thus, it is more accurate in identifying true positive instances. Where the goal is to accurately predict true positive instances from a given set of data. Precision can be calculated in Eq. (1).
$$\varvec{P}\varvec{r}\varvec{e}\varvec{c}\varvec{i}\varvec{s}\varvec{i}\varvec{o}\varvec{n}= \frac{\varvec{T}\varvec{P}}{\varvec{T}\varvec{P}+\varvec{F}\varvec{P}}$$
1
Recall
Recall is defining the ability of a model of correctly identify the all-positive predication over the total actual positive prediction, that called the sensitivity. Which quantifies the proportion of true positive instance out of the total actual positive predication [34]. A highest sensitivity value indicates some false negatives. That means the model is making some incorrect negative predictions, so it is more effective in capturing all the positive predication. Recall is commonly used for an evaluation metric in binary classification approaches, where the main goal is accurately identifying all positive instances from a given datasets of heart vessels, such as detecting diseases, anomalies, or rare events. Recall value calculated in Eq. (2)
$$\varvec{R}\varvec{e}\varvec{c}\varvec{a}\varvec{l}\varvec{l}= \frac{\varvec{T}\varvec{P}}{\varvec{T}\varvec{P}+\varvec{F}\varvec{N}}$$
2
Mean Average Precision
mAP is a values on different recall levels, that provides the overall measurement of all predicated model. All these predicted models is performing in accuracy of object detection. mAP is basically used for metric evaluate the object detection tasks to assess the better performance object detection models with interest of high precision across different recall values[35]. mAP can be calculated in Eq. (3)
$$\varvec{m}\varvec{A}\varvec{P}\left(\varvec{m}\varvec{e}\varvec{a}\varvec{n} \varvec{a}\varvec{v}\varvec{e}\varvec{r}\varvec{a}\varvec{g}\varvec{e} \varvec{p}\varvec{r}\varvec{e}\varvec{c}\varvec{i}\varvec{s}\varvec{i}\varvec{o}\varvec{n}\right)=\left(\frac{1}{\varvec{N}}\right)\sum \left(\varvec{P}\varvec{r}\varvec{e}\varvec{c}\varvec{i}\varvec{s}\varvec{i}\varvec{o}\varvec{n} \varvec{a}\varvec{t} \varvec{e}\varvec{a}\varvec{c}\varvec{h} \varvec{r}\varvec{e}\varvec{c}\varvec{a}\varvec{l}\varvec{l} \varvec{v}\varvec{a}\varvec{l}\varvec{u}\varvec{e}\right)$$
3
F1 Score
This is often a compliment relationship between precision and recall. There could be cases depending on the domain where we would want either precision or recall to be an important metric values [36]. However, generally, we would be create a model that can perform better results on both. Where the F1 score of metric comes in Eq. 4 indicate the F1 score with the help of precision and recall into a single metric.
$$\varvec{F}1 \varvec{S}\varvec{c}\varvec{o}\varvec{r}\varvec{e}= \frac{2\varvec{*}\varvec{P}\varvec{*}\varvec{R}}{\varvec{P}+\varvec{R}}$$
4
Experimental Results:
The proposed approach of deep learning framework for accurate detection of coronary artery blockage using angiography images. The experiment result evaluated using Eq. 1, 2, 3 and 4. The obtained value of precision is 99.4%, a recall is 100%, a mean average precision (mAP) is 99.5%. The value of F1 score of has been calculated as 99.7% with the help of precision and recall using Eq. 4. Precision represents the proportion of true positive instance among all the positive instance made by the model, while recall represents the proportion of true positive predictions among all the actual positive cases in the dataset. mAP is a commonly used metric to evaluate object detection tasks, and it measures the average precision across different levels of confidence thresholds.
Table 1
Performance of Confusion Metrics Parameter Values
Performance Metrics | Predicted Values |
Precision | 99.4 |
Recall | 100 |
mAP | 99.5 |
F1 Score | 99.7 |
From above Table 1 provides the detail information of performance of confusion matric of YOLOv8 model of detection of blockage of coronary artery using angiography images with analysis of precision, recall and mean average precision used evaluation parameters respectively. The graphical representation of evaluation metrics has been shown in Fig. 2.
All these evaluated performances also be present in graphical view with relative confusion metrics parameters in below.
Figure 3 shows that all training and validation losses is reduces by increasing the number of iterations. All reduces parameter is classification loss and box losses with respect to training and validation. The achieved precision of 99.4% indicates a relatively high level of accuracy in identifying the blockage of coronary artery cases as positive by the model. Similarly, the recall 100% indicates that the model is able to detect a significant proportion of coronary artery stenosis cases in the angiography images. Similarly the evaluated results of mean average precision (mAP) is 99.5% indicates an overall good performance in detecting and localizing coronary artery blockages in the images. With the help of these evaluation parameters, the blockage of coronary arteries has been detected with its accuracy. The three sample of original images and blockage detected image with confidence value (accuracy) has been shown in Fig. 4. which is mention in blow.
The confidence detection is a crucial parameter for blockage detection in heart angiography images. Where, a higher level of confidence values provides the better detection of blockages or narrowing of blood vessels. The tested result of image1 having 88% of confidence value, image2 includes 76% of confidence value and similarly image3 having 87% confidence detection in heart angiography images.