1 Introduction

Cardiovascular diseases (CVDs) are indeed the leading cause of death globally. From 1990 to 2019, the number of people affected by these diseases rise of almost 100%, increasing from 271 to 523 million. During the same period, there was a marked rise in deaths due to cardiovascular diseases, rising from 12.1 million cases to 18.6 million cases. Furthermore, disabilities caused by these diseases have doubled, from 17.7 million to 34.4 million. Among the many vascular diseases, artery stenosis stands out as one of the most [1].

Vascular stenosis is a condition that happens when arteries are narrowed due to multiple reasons such as a fatty or blood clots staying in the inner wall of blood vessel. In case of late treatment of this condition may cause severe complications like heart attacks, strokes, or peripheral arterial disease. ML has demonstrated a high ability of improving the accuracy of the diagnostic in medical imaging techniques like CTA [2]. This combination of AI and medical imaging technologies could revolutionize the diagnostic landscape, reducing diagnostic time and enabling immediate interventions. Early detection and management of vascular stenosis are crucial for preventing severe cardiovascular events. Invasive methods like coronary angiography, which carry risks like bleeding, infection, and radiation exposure, are often used for detecting stenosis [3]. Non-invasive methods like CT scans, MRI, and ultrasound are emerging as alternatives. MRI offers a range of insights, but its excessive cost and time-intensive nature may limit its utility in certain medical settings. Ultrasound, on the other hand, has limitations, such as its limited ability to penetrate deeply and its reliance on patient positioning and operator skill [4].

CTA scan analysis is a less invasive and more cost-effective alternative for identifying arterial stenosis and analyzing plaque composition [5, 6]. AI technologies have demonstrated remarkable promise in CTA scan analysis, autonomously detecting and categorizing various anomalies, enhancing the accuracy, efficiency, and speed of CTA scan interpretations [7]. However, manual analysis of CTA images is laborious and time-consuming, particularly in busy hospital settings dealing with high volumes of patient scans. As dependency on CTA scans grows, it is essential to seek innovative solutions that streamline the analysis process, alleviate medical professionals’ pressure, and ensure maintained accuracy in diagnostics [8].

Despite the wide range of articles which show benefits of AI in medical imaging, there is still a lack of comprehensive and systematic reviews that evaluate the accuracy and effectiveness of ML applications in CTA in detecting artery stenosis. The aim of this systematic review is to investigate the following question: In patients undergoing CTA for suspected vessel stenosis or occlusion, how does the accuracy and effectiveness of ML algorithms compare to traditional diagnostic methods, for detecting the severity of artery stenosis? The purpose of this review is to assess the level of knowledge and provide an overview of the relevant literature to guide future research in this field. While there have been various reviews focusing on the integration of AI with medical imaging, in the domain of cardiovascular disease diagnostics [9,10,11,12,13,14,15,16,17,18], there is a significant gap in the existing literature when it comes to the specific use of ML techniques in the detection of anatomical stenosis through CTA.

Based on our knowledge, this is an initial comprehensive evaluation that discusses the utilization of ML techniques for anatomical stenosis detection using CTA. Our review’s goal is to provide new information that will set the standard for future research.

2 Methodology

This systematic review was conducted in accordance with the recommendations provided by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Reviews (PRISMA-DTA) [19]. This systematic review’s protocol has been registered on PROSPERO under the reference number CRD42023460300.

2.1 Inclusion and exclusion criteria

Articles included were research papers and reviews from January 2013 to October 2023, sourced from Web of Science, PubMed, IEEE, and Scopus. They had to be in English and primarily discuss anatomical stenosis detection in arteries via CTA using ML techniques, specifically addressing anatomical stenosis and not functional ones. Exclusions encompassed conference abstracts, letters, editorials, case reports, and studies focusing on multiphase CTA or dynamic CTA. Papers not highlighting ML role in anatomical stenosis detection on CTA, those not in English, multimodality studies used different medical imaging have been excluded as this work is targeting articles that conducted CTA scan images only, types and studies that focus on pulmonary embolism (PE) are excluded too, because PE is more associated with venous thromboembolism than arterial occlusion.

2.2 Retrieval strategy

We scanned PubMed, IEEE Xplore, Scopus, and the Web of Science systematically for literature available from January 2013 until October 2023. Key search terms included “Vascular,” “stenosis,” “Computed Tomography Angiography,” and “machine learning,” along with their synonyms and related terms. Supplemental table S1 provides a full search keyword for each database.

2.3 Literature screening and data extraction

In the initial phase of our systematic review, we collated articles sourced from four distinct databases, meticulously amalgamating the data into a unified Excel sheet while ensuring the exclusion of any duplicates. Subsequently, a pair of reviewers embarked on a screening of the titles and abstracts, aiming to isolate studies that aligned with our predefined inclusion criteria. This process was followed by a detailed evaluation of the full texts of these preliminarily selected studies, determining their relevance and suitability for our research objectives. If a disagreement of opinion arises between the first and second reviewers, it will be resolved through discussions with the third and fourth reviewers. In addition, when needed, some medical data were analyzed, discussed, and evaluated by these additional reviewers. A standardized spreadsheet was used to evaluate the studies, allowing data and results to be extracted in an integrated manner for evaluation purposes.

2.4 Risk of bias assessment and data synthesis

The QUADAS-2 tool was used to assess bias risk and study applicability, accurately covering factors like patient selection, index test, reference standard, and study flow and timing [20]. Our data synthesis approach was meticulously designed to assess the effectiveness and accuracy of ML applications in the detection of vascular stenosis using CTA. Initially, we compiled and standardized the data extracted from the selected studies into an Excel spreadsheet. This facilitated a uniform and systematic evaluation of the collected metrics. For a comprehensive and comparative analysis, the following data items were tabulated: (I) author, (ii) publication year, (iii) study design (retrospective or prospective), (iv) index test (software or ML-based model), (v) number of patients included, (vi) anatomical region of stenosis (coronary, head and neck, peripheral arteries), (vii) reference standard (expert readers or invasive methods), (viii) limitations identified in each study, (ix) stenosis detection threshold, (x) level of evaluation (per patient, vessel, segment, lesion), (xi) all available metrics such as sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), accuracy, and area under the curve (AUC). We utilized Microsoft Excel to systematize these data into a structured format, enabling us to discern patterns and discrepancies across the studies efficiently. This qualitative synthesis allowed us to assess the heterogeneity of the studies based on the variations in stenosis region, problem classification, and evaluation level, which are critical factors in understanding the performance of research conducted in this field.

3 Results

3.1 Literature screening results

A total of 1950 records were identified using the search strategy across multiple databases: IEEE Explore, Web of Science, PubMed, and Scopus. Of these, 1039 duplicates and 103 review papers were removed, leading to 808 records for the initial screening. After reviewing the titles and abstracts, 521 records were excluded, leaving 287 articles for full-text screening. Out of these, 224 reports were excluded due to several reasons: presence of functional stenosis, not related to ML, being irrelevant, utilization of different medical imaging types, dealing with duration, and instances of pulmonary embolism. The flowchart of the PRISMA method used in this study is displayed in Fig. 1.

Fig. 1
figure 1

Flowchart of the PRISMA method, depicting the stages of systematic review, such as identification, screening, determining eligibility, and inclusion

Full size image

Out of the initial selection, 63 studies met our criteria and were subsequently incorporated into our review for a comprehensive qualitative appraisal [5, 8, 21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81]. Figure 2 presents a taxonomy of ML research as it relates to the detection of anatomical stenosis in CTA scans. Our taxonomy shows that most studies focus on coronary artery stenosis, 34 of these studies concentrated on the coronary arteries stenosis detection [5, 8, 50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81]. These were further subdivided, with 15 studies evaluating software tools [53, 59, 61, 70,71,72,73,74,75,76,77,78,79,80,81] and 19 introduced ML methodologies [5, 8, 50,51,52, 54,55,56,57,58, 60, 62,63,64,65,66,67,68,69]. For stenosis detection in the head arteries, 28 studies were identified [22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49]. Of these, 23 assessed software tools [27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49], while a mere 5 introduced ML methodologies [22,23,24,25,26]. However, only one study in our compilation proposed a ML algorithm specifically tailored for peripheral artery stenosis detection [21]. This highlights a potentially under-researched area, hinting at an avenue ripe for future exploration.

Fig. 2
figure 2

Taxonomy of research on ML approaches in identifying stenosis through CTA scans, categorized by the type of ML problem addressed and the discernible stenosis detection threshold (SDT) values

Full size image

Figure 3 shows the increasing number of studies over the past decade that have focused on applying machine learning to detect anatomical stenosis in CTA. Initially, the number of studies during the first five years of the last decade remained low but began to increase significantly since 2019. This upward trajectory underscores the growing academic interest in leveraging machine learning to enhance the accuracy of anatomical stenosis diagnosis in CTA scans.

Fig. 3
figure 3

Number of articles per years

Full size image

3.2 Risk of bias within studies

Figure 4 shows the results of the QUADAS-2 tool evaluation. The results are generally good, indicating a high level of methodological rigor in the subject. A significant majority of the studies demonstrated a low risk of bias in crucial areas such as patient selection, index testing, and reference standards. This demonstrates a systematic and dependable methodology in the implementation of ML algorithms and the standards employed for comparisons, highlighting the reliability of these techniques in research settings. Despite the overall low risk in key domains, some variability was noted in the Flow and Timing category, highlighting a potential area for improvement in future research. Moreover, the low applicability concerns across the studies suggest that the findings are not only academically robust but also highly relevant to our review subject. Supplemental Table S2 provides a summary of bias and questions regarding applicability.

Fig. 4
figure 4

The QUADAS-2 tool graph illustrating the potential for bias and applicability of the studies

Full size image

3.3 Coronary artery stenosis detection

Coronary artery stenosis is a condition when the main arteries in the heart are narrowed or obstructive [7]. Coronary CTA has gained prominence as a non-invasive method for diagnosing suspected cases of coronary artery disease. Coronary CTA has proven adept at quantifying the extent of coronary stenosis and providing detailed insights into the morphology and composition of coronary artery plaques. However, the visual assessment of the entire coronary tree to determine the severity of stenosis is subjective, time-consuming, and dependent on the expertise of the physician. Therefore, the development of an accurate and automated method to assist physicians in identifying coronary artery stenosis would be beneficial in clinical practice [69]. The Coronary Artery Disease—Reporting and Data System (CAD-RADS) consensus statement offers a detailed classification system for the severity of stenosis. Under this system, stenosis is categorized as follows: CAD-RADS-0 indicates an absence of stenosis (0%), CAD-RADS-1 denotes minimal stenosis ranging from 1 to 24%, CAD-RADS-2 is used for mild stenosis (25–49%), CAD-RADS-3 represents moderate stenosis (50–69%), CAD-RADS-4 signifies severe stenosis (70–99%), and CAD-RADS-5 corresponds to a total occlusion, implying 100% stenosis [76].

The proficiency of ML tools in detecting coronary artery stenosis demonstrates remarkable versatility, adapting to various anatomical scales. At the patient level, these algorithms undertake a comprehensive assessment, gauging the presence or absence of stenosis across the entire coronary network. At the vessel level, ML algorithms focus on each artery separately, tracking its entire path in order to identify any instances of stenosis. At the segment level, ML algorithms divide each artery into multiple segments and scan each artery segment separately to check for stenosis. Locating a stenosis in a segment of an artery is a more difficult and complex job. At the lesion level, where each stenotic lesion is meticulously analyzed by ML algorithms, parsing through the unique characteristics that define it. In Tables 1 and 2 provide a comprehensive overview of various studies, evaluating the effectiveness of ML-based software and algorithms in detecting arterial stenosis and occlusion using Coronary CTA scans. The results are evaluated using classification metrics, which take a problem of stenosis detection as classification problem (presence or absence of stenosis) in Coronary CTA scans per patient, per vessel, per segment, or per lesion. However, there are some studies that used quantitative metrics for stenosis detection described in appendices Table 3.

Table 1 An overview of research examining the efficacy of ML-based software in identifying arterial stenosis through coronary CT angiography scans where InCA is invasive coronary angiography, FFR is fractional flow reserve, QCA is quantitative coronary angiography, SE is sensitivity, SP is specificity, PPV is positive predictive value, NPV is negative predictive value, ACC is accuracy, AUC is area under the curve, DSA is digital subtraction angiography, and __ symbol means not recorded value
Full size table
Table 2 An overview of research examining the efficacy of ML algorithm in identifying arterial stenosis through coronary CT angiography scans, where MCC is Matthews correlation coefficient, SE is sensitivity, SP is specificity, PPV is positive predictive value, NPV is negative predictive value, ACC is accuracy, AUC is area under the curve, and __ symbol means not recorded value
Full size table
Table 3 An overview of research examining the efficacy of ML algorithm in identifying arterial stenosis through coronary CT angiography using quantitative metrics and __ symbol means not recorded value
Full size table

The results depend on several parameters, such as the anatomical levels of stenosis classification, the complexity of the classification (whether it is binary or multiclass), and the detectable SDT value. The ML algorithms vary with the scale of the analysis. The broader scales, like patient and vessel levels, show more effective performance, but the algorithms encounter more obstacles as they examine the finer details, such as segments and lesion levels. This variation underscores the potential for further refinement and development, particularly in analyses at the segment and lesion levels. These observations are vital for steering the direction of future research and advancements in the field. At the patient level, we have observed high sensitivity and specificity, clearly showcasing the algorithms’ adeptness at correctly detecting the stenosis. We can notice a variability in positive predictive value and negative predictive value values across studies, likely due to variations in algorithmic precision and varying SDT. While these metrics generally have high performance, they often dip in sensitivity and specificity at the vessel level compared to patient-level outcomes. The segment level also shows a decrease in sensitivity and specificity, indicating difficulty in detecting stenosis in specific arteries or artery segments. Most studies in Tables 1 and 2 use SDT ≥ 50%. While the performance of these studies tends to improve at higher SDT, like ≥ 70%, it is because the symptoms of severe stenosis are more evident. Additionally, the use of the CAD-RADS system, despite offering a comprehensive assessment framework, shows varied effectiveness and tends to have lower performance due to its complexity, especially in the automated detection of stenosis. Binary classification studies, which detect stenosis based on a single SDT value, are prevalent and demonstrate powerful performance across various anatomical levels. Multiclass classification studies provide a more detailed understanding of stenosis severity but present challenges in their complexity and implementation.

On the left side of Fig. 5 we show the number of studies according to the level of stenosis detection threshold (SDT) used. The graph on the left indicates that most studies have focused on detecting stenosis of 50% or more, most likely because this level of stenosis often requires treatment. On the right side of Fig. 5, we clarify the difference between the number of studies that detect stenosis according to the study’s ability to classify between types of coronary artery stenosis, binary classification, and multiclass classification. This indicates that detecting the presence or absence of stenosis is a priority in current studies, while there is a clear lack of studies that can detect different degrees of stenosis in the same model.

Fig. 5
figure 5

Number of studies per classification type and SDT, where BC is binary class classification and MC is multi-class classification

Full size image

In Tables 2 and 3 we can find a variety of ML algorithms to identify arterial stenosis through coronary CT. Early studies [52, 55, 67, 68] employed image processing and feature extraction methods, followed by ML methods for stenosis detection. Studies [50, 54] use the ConvLSTM model in plaque segmentation. This task entails recognizing different components of the plaque which contribute to stenosis severity upon successful classification. Study [57] proposed M-Net for artery segmentation, followed by automatically calculate MLA, DS, and CDD via CNN. Similarly, study [58] first used a U-Net model for artery segmentation to find ROI before applying a 3DNet model for classification. Study [5] introduced a Mask-RCNN model followed by GBDT after stenosis classification where U-Net was chosen for artery segmentation and ROI detection. A deep CNN-based model for stenosis classification at patient level from CTA images is presented in study [65]. Multiplanar reconstruction (MPR) is used as an input data in studies [62, 64, 66, 69]. MPR is a medical imaging technique and uses CTA scans to produce two-dimensional images from three-dimensional data. Study [69] proposed ConvMixer architecture, a deep learning architecture tailored for coronary stenosis classification based on CAD-RADS. Similarly, study [62] proposed TR-Net, study [64] proposed 2.5D CNN, and study [66] used the RCCN for stenosis detection. Curved Multiplanar Reformation (cMPR) is used as input data in studies [8, 51, 60]. cMPR is an expansion of the MPR technique that allows the imaging of a specified curved plane along a centerline of a vessel, identifying the vessel’s complete line in a single two-dimensional image. In study [51], the others used the InceptionV3 architecture for stenosis classification. Similarly, study [60] utilizes an attention-based architecture incorporating the squeeze-and-excitation network module and study [8] proposed the coronary R-CNN method for stenosis detection for stenosis detection.

3.4 Head and neck arteries stenosis detection

The stenosis in head and neck arteries can increase stroke risk by reducing blood flow to the brain. Large Vessel Occlusion (LVO) is a blockage in major arteries crucial for brain circulation, including the internal carotid artery ICA, middle cerebral artery MCA, anterior cerebral artery ACA, posterior cerebral artery PCA, vertebral artery VA, and basilar artery BA. Blockages in these arteries can have significant implications for brain health and function. Acute Ischemic Strokes (AIS) are often caused by LVOs, leading to high morbidity and mortality [77]. Patients suffering from ischemic stroke (AIS) can be saved if they are treated within 24 h of the start of symptoms using endovascular thrombectomy (EVT) [78]. The critical influence of EVT on patient outcomes depends on the prompt and accurate detection of an LVO. Non-invasive imaging techniques like CTA or MRI are crucial for identifying LVOs. ML-based models and software tools in the realm of head and neck CTA imaging can be broadly categorized into two main types: occlusion detection and stenosis detection. The evaluation of these studies is based on various anatomical analysis levels, evaluations at patient, vessel and segment level and includes all relevant vessels in the head and neck area.

Tables 4 and 5 provide an overview of studies evaluating the effectiveness of ML-based software and algorithms in detecting arterial stenosis and occlusion using CTA in head and neck arteries. The results are evaluated using classification metrics, which take the problem of stenosis detection as a classification problem (presence or absence of stenosis) in CTA scans per patient, per vessel, per segment, or per region. Furthermore, in some studies focused on occlusion detection, the evaluation criteria are often tailored to specific vessels. This includes assessing occlusions in the ICA and MCA-M1 Segment (lLVO). Expanding this scope, some studies also examine occlusions in the ICA, MCA-M1, and MCA-M2 (eLVO). Furthermore, the most comprehensive of these studies extends their analysis to include occlusions in the ICA, MCA-M1, MCA-M2 segments, and other proximal intracranial vessels (pLVO). However, there is one study that deals with the stenosis localization problem and uses object detection metrics as described in appendices, Table 6.

Table 4 An overview of research examining the efficacy of ML-based software in identifying arterial stenosis through CT angiography scans of head and neck arteries, where SE is sensitivity, SP is specificity, PPV is positive predictive value, NPV is negative predictive value, ACC is accuracy, AUC is area under the curve, and __ symbol means not recorded value
Full size table
Table 5 An overview of research examining the efficacy of ML models in identifying arterial occlusion through CT angiography scans in head and neck arteries, where SE is sensitivity, SP is specificity, PPV is positive predictive value, NPV is negative predictive value, ACC is accuracy, AUC is area under the curve, and __ symbol means not recorded value
Full size table
Table 6 An overview of research examining the efficacy of ML models in localizing arterial occlusion through CT angiography scans
Full size table

On the left side of Fig. 6 we show the number of studies according to the level of stenosis detection threshold (SDT) used. The graph on the left indicates that most studies have focused on detecting LVO. On the right side of Fig. 6, we clarify the difference between the number of studies that detect stenosis according to the study’s ability to classify between types of head and neck artery occlusion and stenosis, binary classification, and multiclass classification. This indicates that detecting the presence or absence of occlusion is a priority in current studies, while there is a clear lack of studies that can detect different degrees of stenosis in the same model.

Fig. 6
figure 6

Number of studies per classification type and SDT, where BC is binary class classification and MC is multi-class classification

Full size image

Most studies deal with vessel occlusion detection in head and neck CTA scans. The studies demonstrate that ML tools are particularly effective in detecting eLVO and lLVO. These tools provide reliable results in diagnosing these occlusions, which is critical for stroke diagnosis and treatment planning. The performance in pLVO detection has lower sensitivity, highlighting a potential area for further research and development.

We found two studies in the literature dealing with stenosis detection at head and neck arteries. The two studies evaluate the performance of CerebralDoc software [42, 45]. CerebralDoc software exhibits powerful performance across all anatomical levels, with its ability slightly diminishing as the scale gets finer, which is common in medical imaging due to increased complexity at finer scales. The software shows particularly high specificity across all levels, ensuring the reliable exclusion of false positives. In Tables 5 and 6, we find a limited number of studies that used ML algorithms to identify LVO in head and neck arteries through CTA scans. Studies [22,23,24, 26] proposed deep learning models for image segmentation and LVO detection. Study [25] employs DL model for improving the detection of cerebral arterial stenosis in general using a retrained two-stage detection algorithm. The study proposes a methodology to localize stenosis but does not delve into quantifying or grading the severity of the stenosis.

The studies in Table 4 focus on a SDT equal to 100%, reflecting the clinical importance of identifying arterial occlusion. While CAD-RADS is a coronary artery specific system, its principles of grading stenosis severity provide a useful reference for understanding the performance of ML tools in head and neck artery analysis. The studies in Tables 4 and 5 effectively demonstrate the capabilities of ML tools in detecting severe stenosis and occlusions in head and neck arteries, aligning with the higher grades of the CAD-RADS system. The study highlights the importance of identifying severe stenosis and occlusions in stroke management and neurological diagnostics for effective treatment. However, there is a lack of research on lower grades of stenosis, indicating a need for further development.

3.5 Peripheral arteries stenosis detection

Peripheral artery stenosis, also known as Peripheral Artery Disease (PAD), is a serious medical condition where the arteries supplying blood to the arms and legs become narrow or clogged with plaque. This can reduce blood flow, causing pain and cramping symptoms including numbness, or weakness in the involved organs. Additionally, the condition can raise health risks serious conditions such as heart disease, stroke, and gangrene have increased [21]. The proposed study [21] examines peripheral artery stenosis; data from 265 patients were meticulously analyzed using the p-EffNet, a convolutional neural network (CNN), to evaluate arterial stenosis and occlusion in lower extremities. This study delineated stenosis into various grades above knee (0%, 1%–50%,51%–69%,70%–99%, 100%) and below knee (0%–50%,51%–99%,100%), achieving a sensitivity of 90.2%, specificity of 97.7%, and accuracy 91.5% for above-knee arteries, and a sensitivity of 91.3%, specificity of 95.2%, and accuracy 90.9% for below-knee arteries. For below-knee arteries, the classification is less granular than for above-knee arteries due to the smaller diameter of these vessels, which makes precise assessment of stenosis percentages more challenging with CTA.

3.6 Evaluation metrics

Guided by [82] we conducted a comprehensive search to find all metrics used in evaluating the artery stenosis detection problem. In all studies [5, 8, 11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70], a wide array of metrics has been employed to evaluate the performance of various algorithms and techniques. Figure 7 shows the distribution of frequently used evaluation metrics in stenosis detection per studies. Classification metrics such as sensitivity, specificity, PPV, NPV, accuracy, AUC, MAE, AUC-ROC, F1Score, and MCC are prominently used, with sensitivity being the most frequently reported metric. This suggests a significant focus on the ability of models to correctly identify true positives, which is vital in medical diagnostics to avoid missing any cases of stenosis. Quantitative metrics like Minimal Lumen Area, Diameter Stenosis, and Contrast Density Difference are used less frequently than classification metrics. This could imply that while quantitative analysis is important, the primary focus of these studies might be on the classification performance of the models. The Object Detection Metric, mAP of IoU, also shows a lower frequency of use compared to the classification metrics. The infrequent use of the object detection metric mAP of IoU indeed suggests a potential gap in the literature concerning the localization of stenosis.

Fig. 7
figure 7

Distribution of frequently used evaluation metrics in stenosis detection per study

Full size image

4 Discussion

Cardiovascular diseases have become the leading cause of death globally; one of the most common vascular diseases is stenosis. Using CTA scans, radiologists and cardiovascular doctors can evaluate plaque location and luminal stenosis. Manual CTA scan analysis is laborious and time-consuming. ML algorithms have shown a high ability to detect arterial stenosis in CTA scans. This research involved a thorough examination of 63 studies, and the resulting findings are as follows:

4.1 Coronary artery stenosis detection

ML algorithms are effective in stenosis detection, particularly at broader anatomical levels. There is a notable trend of increased difficulty and decreased performance metrics at finer levels of analysis. Most studies focus on a binary classification approach with a common SDT of ≥ 50%, aligning with clinical needs for significant stenosis detection. A primary research gap lies in the prevalent use of binary classification models, which simplify the detection into a basic ‘presence or absence’ of stenosis. This approach, while useful for certain clinical decisions, fails to capture the smaller stenosis severity. The challenge lies in effectively developing and implementing ML models that can accurately perform multiclass classification. These models need to distinguish between multiple levels of stenosis severity, which is inherently more complex than binary classification. Another notable gap is the limited focus on detecting mild stenosis (which represents early-stage disease), an early-stage disease where intervention can be most beneficial. The challenge in detecting mild stenosis lies in the fact that ML algorithms need to be highly sensitive to small changes in the arteries, which requires high-quality imaging data and advanced analytical capabilities. Moreover, distinguishing mild stenosis from normal variations in arterial anatomy can be difficult. Additionally, there is a lack of emphasis on the precise localization of stenosis within the arterial system, which is critical for treatment planning. Accurately localizing stenosis is challenging due to the complex and three-dimensional nature of arterial structures. The challenge is compounded when considering the three-dimensional nature of arterial structures and the need for models to differentiate between clinically significant and insignificant findings.

4.2 Head and neck artery stenosis detection

ML tools are particularly effective in identifying extensive large vessel occlusion (eLVO) and large vessel occlusion (lLVO). This capability is crucial for stroke diagnosis and treatment planning, where timely identification of these occlusions can significantly impact patient outcomes. Most studies deal with detecting arterial occlusions (SDT equal to 100%). While these studies are effective in identifying significant occlusions, this approach may miss milder forms of stenosis. The challenge lies in extending the capability of ML tools to detect and quantify lower grades of stenosis, which are equally important for early intervention and preventive strategies. The performance of tools like CerebralDoc in detecting finer-scale stenosis (such as pLVO) seems to exhibit lower sensitivity. Enhancing sensitivity at finer anatomical levels requires sophisticated image analysis techniques. The complexity increases at these finer scales, demanding more advanced ML algorithms and higher-quality imaging data. Although there is one study that proposes methodologies for localizing stenosis, the development of ML models that can detect and quantify stenosis severity is crucial.

4.3 Peripheral artery stenosis detection

Peripheral artery stenosis detection in CTA scans using ML-based models presents a promising and largely untapped area of research within medical imaging and diagnostics. The singular study using the p-EffNet convolutional neural network to evaluate stenosis in lower extremities demonstrates the potential efficacy of such tools in diagnosing this condition. However, the fact that only one study has been conducted so far highlights a significant gap in the field. This scarcity of research underscores the immense opportunity for further exploration and development in applying ML algorithms to peripheral artery stenosis. Given the clinical importance of early and accurate detection of this condition, further studies could greatly contribute to advancements in treatment planning and patient care, making it an exciting and valuable subject for future research endeavors.

4.4 Included study limitations

In included study several common limitations have emerged, highlighting critical areas for future improvement. A notable limitation is the retrospective nature of many studies [5, 8, 21, 24, 25, 27, 28, 30,31,32, 34,35,36,37,38,39,40,41,42,43,44,45,46,47, 49, 51,52,53, 55, 58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81], which can introduce biases and limit the ability to gauge real-time impacts on clinical decision-making. The generalizability of findings is often constrained by single-center designs or limited datasets, which may not accurately represent broader patient populations [21, 24, 26, 34, 35, 37, 40, 41, 46, 47, 53, 55, 57,58,59,60,61,62,63,64,65,66, 69, 72, 74, 75, 79]. Some studies are interpreted by only one expert, typically without additional validation. This approach can introduce bias, as the findings are subject to the individual’s perspective, expertise, and potential errors [22, 40, 41, 45, 48, 49, 60]. Technical limitations, such as the exclusive use of specific types of scanners or Technical Difficulties with Data Transmission, further restrict the applicability of these studies [26, 35, 51, 53, 55, 60, 63, 66, 69, 79]. The studies study’s performance and evaluation may be affected by the limited number of patients included in each study [45, 55, 59, 64, 68, 81]. Furthermore, the use of ML in medical imaging requires standardized and robust validation methods to address variability and potential bias in expert readings, ensuring accuracy, reliability, and effectiveness across diverse clinical settings.

4.5 Future research

A comprehensive analysis of 63 studies on the use of ML in CTA scans for detecting stenosis highlights several key areas for future research. These recommendations aim to address current gaps and improve the effectiveness of ML applications in medical imaging, particularly in the context of cardiovascular diseases. Future studies should focus on developing and refining machine learning algorithms for efficient multiclass classification of stenosis severity, enabling more nuanced understanding and personalized patient treatment strategies. Research is needed to improve stenosis localization in the vascular network, aiming to create machine learning models that accurately map and characterize stenosis in a three-dimensional space. Research on stenosis detection in head and neck arteries using machine learning is limited, but expanding this area is crucial for stroke diagnosis and treatment, as early identification can significantly impact stroke patient management and outcomes. The current ML tools’ detection capabilities for small-scale stenosis, like pLVO, are limited and need more studies to enhance detection capabilities. Public datasets are essential for training robust machine learning models, encompassing a wide range of stenosis cases. Encouraging the publication and sharing of high-quality imaging data can facilitate more extensive and inclusive research in this field. The limited research in peripheral artery stenosis detection using ML indicates a significant area for future investigation. Given its clinical importance, more studies in this area could lead to substantial advancements in patient care.

4.6 Study limitations

In this study, many limitations should be acknowledged to contextualize the findings accurately. Firstly, the review’s temporal scope is limited to studies published within the last ten years. Additionally, restricting the review to articles published exclusively in English could introduce a language bias, potentially overlooking significant contributions from non-English sources. The reliance on 63 selected studies also presents a challenge, as these studies may vary in terms of design, methodology, and demographic focus, thereby impacting the generalizability of the review’s conclusions.

5 Conclusion

ML’s integration into CTA for vascular stenosis detection represents a significant stride forward in cardiovascular diagnostics. While the current landscape shows promising results, particularly in coronary and head and neck arteries, there remains substantial scope for development, especially in algorithm refinement, stenosis localization, and extending research to peripheral artery stenosis. Addressing these gaps will not only advance scientific understanding but also translate into improved clinical outcomes in diagnosing and managing vascular diseases.