One of the world’s greatest contributors to mortality and morbidity is cardiovascular disease (CVD), which accounts for about 18 million deaths per year [1]. The primary two causes of CVD-related fatalities are coronary artery disease (CAD) and acute coronary syndrome (ACS) [2]. Generally speaking, CAD entails the shrinking of arteries as a result of the buildup of atherosclerotic plaque within their walls, resulting in coronary artery obstruction [3]. Aiming to enhance the diagnosis and treatment of heart disorders as well as lowering the fatality rate from CVD, significant advancements have been made in cardiovascular research and therapy in recent decades [4]. It is now possible to carry out a comprehensive qualitative and quantitative assessment of heart morphological structures as well as operations with the use of contemporary medical imaging techniques, including intravascular ultrasound (IVUS) ^{[5][6][7][8][9]}[5,6,7,8,9], computed tomography (CT) [10], magnetic resonance imaging (MRI) ^[11][12][13][11,12,13], and ultrasound (US) ^[14][15][14,15], which assist identification, disease monitoring, surgical planning, and evaluation. An example of the coronary artery is shown in Figure 1a, while the IVUS acquisition device for the coronary vascular system is shown in Figure 1b.

The diagnosis of CAD is frequently made by coronary CT angiography (CCTA), which enables non-invasive measurement of the arterial lumen’s diameter and plaque localization ^{[16][17][18][19][20]}[16,17,18,19,20]. However, radiologists presently manually assess the location and severity of the plaque(s) leading to the stenosis in CCTA pictures, which, in addition to being costly and time-consuming, is also susceptible to mistake and inaccuracy [21]. In order to develop computerized and accurate coronary artery stenosis as well as a plaque identification method, it is crucial that coronary arteries in CCTA pictures must be automatically segmented. The following factors, however, make automatic coronary artery segmentation for CCTA pictures particularly complicated. To begin with, coronary circulation has a complicated pattern, with several arteries of different thicknesses [22]. For perfect segmentation, some of the branches are even too thin. Additionally, individual differences in the structure of the coronary artery tree may be relevant. Second, other vascular organs that seem like the coronary arteries adjacent to the heart can be mistaken for them because of their similar appearance [23]. Third, the coronary arteries only make up a tiny fraction of the entire heart’s cells, and the methods for segmentation must consider this imbalance [24]. Additionally, several variables, including heart rate, the data reconstruction method, the quantity of the injected contrast agent, and radiation exposure, affect the quality of the pictures obtained during CT angiography [25]. Coronary artery segmentation, therefore, is more challenging due to low-resolution image quality.

Figure 1b shows a popular imaging method for the assessment and control of CVD, intravascular ultrasound (IVUS) ^[5][26][5,26]. In conjunction with positional data, IVUS images are segmented into interior and exterior regions as lumen and media regions, respectively. Arteries’ representation in 3D heavily depends on arterial vessel walls for various purposes such as surgical planning. The arteries’ segmentation is helpful for plaque identification in clinical practices. IVUS-guided percutaneous coronary intervention (PCI) is a more advanced and superior technique in comparison to standard angiography-guided PCI, minimizing death risks in patients [6]. IVUS segmentation for lumen and vessel cross-sectional based on 3D vessel reconstruction is precise and quick for accurate and real-time segmentation during PCI [27]. However, IVUS segmentation requires recent, accurate, and faster techniques, typically at 30 Hz and 100 Hz frame rates. To record an IVUS sequence, a catheter-borne ultrasound transducer is inserted into the coronary artery and then returned via arteries at a speed of roughly 1 mm/s [5]. Raw radio frequency (RF) information from the probe is typically not used for analysis. However, amplified and filtered gray-scale B-mode Euclidean ultrasound pictures showing the coronary cross-section provide a typical output format for downstream evaluation (see Figure 2) The arrows in Figure 2 depict a typical example of five (1–5) frames with calcified plaques. Six patients’ IVUS videos’ worth of frames were collected, and they were placed in a 6 × 5 matrix. The symbol I (1,1)-I (6,5) is used to represent this.

IVUS segmentation is one of the most challenging tasks in medical images. It consists of lumen–intima (LI) and media–adventitia (MA) border detection. This challenge is due to the presence of the artifacts, namely shadows, bifurcation, and echogenic plaques, and the fact that public expert-labeled ground-truth databases only contain a small number of captures [28]. Even though artificial intelligence (AI) has shown promising signs toward higher accuracy and learning strategy, it has been observed that these AI-based black boxes lack clinical validation and the ability to perform well in clinical settings, and they are unable to explain the outcomes ^{[29][30][31][32][33][34][35]}[29,30,31,32,33,34,35]. The clinical validation requires that the outcome from the AI system must have a behavior leading to correct coronary artery disease risk assessment. For example, should an AI system perform accurately on a test patient who has a high risk, then the syntax score of this patient should be high [36]. Other ways to show the clinical validation include by estimating the relationships or correlations between two quantities such as computed tomography (CT) coronary artery score vs. AI outcome of the risk [37]. Such consistent behavior needs to be exhibited by AI systems. Other than the clinical validation, there are attributes such as imbalanced classes in the datasets that can introduce AI bias [38]. Such causes can lead to bias in AI modules or system designs.