Deep Learning Segmentation Techniques for Atherosclerotic Plaque on Ultrasound Imaging: A Systematic Review

Abstract

Background: Atherosclerotic disease is the leading global cause of death, driven by progressive plaque accumulation in the arteries. Ultrasound (US) imaging, both conventional (CUS) and intravascular (IVUS), is crucial for the non-invasive assessment of atherosclerotic plaques. Deep learning (DL) techniques have recently gained attention as tools to improve the accuracy and efficiency of image analysis in this domain. This paper reviews recent advancements in DL-based methods for the segmentation, classification, and quantification of atherosclerotic plaques in US imaging, focusing on their performance, clinical relevance, and translational challenges. Methods: A systematic literature search was conducted in the PubMed, Scopus, and Web of Science databases, following PRISMA guidelines. The review included peer-reviewed original articles published up to 31 January 2025 that applied DL models for plaque segmentation, characterization, and/or quantification in US images. Results: A total of 53 studies were included, with 72% focusing on carotid CUS and 28% on coronary IVUS. DL architectures, such as UNet and attention-based networks, were commonly used, achieving high segmentation accuracy with average Dice similarity coefficients of around 84%. Many models provided reliable quantitative outputs (such as total plaque area, plaque burden, and stenosis severity index) with correlation coefficients often exceeding R = 0.9 compared to manual annotations. Limitations included the scarcity of large, annotated, and publicly available datasets; the lack of external validation; and the limited availability of open-source code. Conclusions: DL-based approaches show considerable promise for advancing atherosclerotic plaque analysis in US imaging. To facilitate broader clinical adoption, future research should prioritize methodological standardization, external validation, data and code sharing, and integrating 3D US technologies.

1. Introduction

Atherosclerotic cardiovascular disease (ASCVD) is the leading global cause of mortality and morbidity, accounting for approximately two-thirds of cardiovascular (CV) disease deaths [1,2]. Underlying the pathogenesis of ASCVD is atherosclerosis, a multifactorial process progressing through different phases of arterial wall lesions, from the early fatty streaks to the development of fibrous plaques. Plaque build-up may lead to stenosis or occlusion of the vessel lumen that, along with plaque rupture or erosion, puts patients at a higher risk of occlusive thrombosis and adverse outcomes, including coronary artery disease, myocardial infarction, and stroke [3]. Unfortunately, in humans, ASCVD evolves over decades and remains asymptomatic until it has already reached an advanced stage. Therefore, the assessment and classification of ASCVD risk is crucial to support clinical decision-making, with enormous prognostic implications [4].

Atherosclerotic plaque imaging relies on various techniques, including computed tomography, magnetic resonance, positron emission tomography, and ultrasound (US). US is routinely applied to image atherosclerotic lesions in multiple arterial territories, including carotids, coronaries, and peripheral vessels [5]. Among the possible applications, conventional carotid US has emerged as a widely available and feasible method for estimating and stratifying the ASCVD risk, particularly when combined with traditional risk factors. In this line, several parameters have been investigated for their association with plaque development, progression, and the likelihood of future adverse CV events. Among these, the assessment of carotid intima–media thickness (IMT) has been highlighted [6,7,8]. IMT refers to the thickening of the intimal and medial layers of the arterial wall and is considered an early biomarker for atherosclerosis risk stratification; in fact, an increased IMT is associated with CV events [9,10]. On the other hand, in coronary applications, intravascular ultrasound (IVUS) is used to estimate key parameters such as the media–adventitia border and lumen area, which are prevalently employed for diagnosing and assessing coronary artery disease. Furthermore, IVUS provides valuable insights into plaque composition in patients undergoing invasive angiography, enabling the differentiation of stable/unstable patterns of atherosclerotic lesions and facilitating the optimization of intervention strategies [5]. Recently, beyond merely assessing luminal stenosis, there has been a focus on the novel analytical applications of atherosclerosis imaging, including ultrasonography, to examine the pathological processes occurring within the plaque. These processes, including inflammation, fibrosis, calcification, and neoangiogenesis, contribute to a comprehensive understanding of plaque phenotypes [11].

US carotid and coronary imaging relies on the robust segmentation of anatomical structures from images to monitor the progression from health to pathology in the arterial vessels, lumen, and plaque. Traditionally, the segmentation of lumen vessels and plaques in US images has been accomplished using classical image processing techniques [12], such as edge detection- and contour-based methods, including the Snakes model [6,13,14], level set method [15], and their combinations [16]. Additionally, machine learning (ML)-based approaches employing clustering methods, like k-means and support vector machine, have been developed and used for vessel and plaque segmentation tasks [17,18]. However, these solutions face significant challenges, primarily concerning image quality issues, including poor contrast, brightness inhomogeneities, anatomical variability, and operator dependence.

To address these limitations, semi- and fully automated analysis of vascular US images has been rapidly implemented, leveraging artificial intelligence (AI) techniques—particularly deep learning (DL)—to improve precision, minimize inter-operator variability, and reduce the time required for analysis. DL-based image segmentation, which utilizes complex neural networks for pattern recognition from images, has become increasingly important, providing an end-to-end solution and minimizing the need for human intervention or predefined parameters [19]. This has significantly advanced the quantitative analysis of plaque parameters [20], enabling the assessment of lumen stenosis at the site of atheroma, plaque burden (i.e., plaque content, volume, and distribution) [21,22], as well as plaque composition and remodeling [23]. Notably, direct plaque segmentation is more challenging than vessel segmentation due to the heterogeneity of plaque composition and shape, and the poor contrast of plaques in US imaging. Currently, these tasks are tackled using hybrid approaches that combine vessel segmentation with plaque detection and/or characterization. Moreover, annotated medical datasets are often small and are time-consuming to acquire, leading to the frequent use of transfer learning (TL) techniques [24]. TL involves leveraging DL models that have been pre-trained on large datasets, which are then fine-tuned for specific or related/similar tasks. Common architectures used in this context include Unet [25], DeepLab [26], SegNet [27], PSPNet [28], DeepLab V3+ [29], and transformer-based models [30].

Despite the potential of DL methods to enhance image segmentation and enable rapid and unbiased analyses, exploring their limitations and establishing trustworthiness for broader adoption in medical image analysis is essential. This area is rapidly emerging as a significant focus of research.

In this overview, we examine recent DL-based algorithms for image segmentation in atherosclerotic plaque US imaging. Our focus is on DL methods applied to conventional US (CUS) and IVUS, which are the primary imaging techniques for segmenting, characterizing, and classifying atherosclerotic plaques in carotid and coronary vessels, respectively. After reviewing recent applications, this review reveals that DL-based methods can effectively manage complex tasks related to medical US image analysis, particularly those concerning plaques, and demonstrate a promising performance based on key task-related metrics. However, improving reproducibility and generalizability is crucial for enhancing the clinical relevance of these methods. The availability of large datasets, better data accessibility, and efforts in external validation could facilitate this advancement. Finally, we discuss the limitations of these methods and the potential future directions for applying DL methods in the context of medical US segmentation.

2. Materials and Methods

This section describes the search strategy adopted and explains in detail all the inclusion/exclusion criteria that led to the collection of the final works in this review.

2.1. Search Strategy

This systematic review was carried out using the PubMed/Medline electronic, Scopus, and Web of Science (WoS) databases. The search was conducted according to the Preferred Reporting for Systematic Reviews and Meta-Analysis (PRISMA) guidelines [31]. The systematic search included studies published up to 31 January 2025 without language restrictions and with a retrospective time limit on publication years.

We performed advanced research, concatenating terms with Boolean operators. In particular, search words and key terms used in the search included (“Segmentation”) AND (“plaque” OR “stenosis”) AND (“deep learning” OR “neural-network” OR “neural networks” OR “CNN” OR “UNet” OR “U-Net” OR “net”) AND (“ultrasound” OR “echography” OR “sonography”). Three authors separately performed the literature search (L.D.R., E.M.S., M.A.).

2.2. Eligibility Criteria

Studies were selected according to the following inclusion/exclusion criteria:

Inclusion criteria

Studies developing/proposing/applying DL algorithms for directly segmenting atherosclerotic plaques on US images or papers using DL architectures to segment the lumen/intima–media or something else in the vessels, but using segmentation outputs to characterize/classify severity/assess geometric features and/or quantitative parameters of plaques in US images.

No particular restriction either on the type of DL architecture used in the studies or on the vascular sites considered was applied. The only restriction on DL models developed/used was that they had to be segmentation networks.

Exclusion criteria

Studies were restricted to peer-reviewed original research articles. Accordingly, the following publication types were excluded: reviews, conference proceedings, conference abstracts, book chapters, letters, and editorials. Studies proposing AI-based segmentation methods, but not deep architectures, were excluded. Publications not in the English language were excluded.

2.3. Data Analysis

Three investigators (L.D.R., M.A., E.M.S.) independently conducted the initial screening of all publications identified through online research by reviewing the titles and abstracts. Any disagreements among the reviewers were resolved by consensus through discussion. Studies published by the same research groups over the years or using the same datasets were included in the review process. After selecting the papers, we collected the following characteristics for each article: sample size (number of patients/images involved in the study), anatomical district investigated, US technology used, study aim, proposed task on plaques, and the main results obtained. The flow diagram of the study selection—according to PRISMA guidelines [31]—is shown in Figure 1. All reports included in this systematic review were analyzed by grouping studies by anatomical district of interest (i.e.,

Table 1. Main characteristics of the 38 original articles included in the analysis and focused on carotid vessel US imaging.

Ref.	N Subjects/N Images	US Imaging	Carotid Site	Aim of the Study	Proposed Task	Main Results
[32]	844/1270	CUS	CCA, ICA, ECA	Plaque segmentation and classification with multi-task learning framework using RCCM-Net	Segmentation, Classification	DSC: 84.92 ± 0.40; R TPA: 0.939; Acc: 91.1/92.3 hyper-/hypo-echoic plaques
[33]	na/450	CUS	na	Plaque segmentation using UNet integrating a self-attention mechanism	Segmentation	DSC: 80.8 ± 15.8
[34]	83/7036	CUS	CCA	SSI assessment and presence/absence plaque classification via portable free-hand 3D-US system	Classification, Quantification	R SSI%: 0.76; Acc plaque yes/no 92/80 frame-/video-based
[35]	844/1270	CUS	na	Plaque segmentation/classification (hypo-/hyper-/mixed-echoic) using an end-to-end multi-task learning network	Segmentation, Classification	Acc: 90.7/92.1 hyper-/hypo-echoic plaques
[36]	88/11,048	CUS	CCA	Quantification of maximum PB on US-video	Quantification	R PB: 0.61
[37]	413/4652	CUS	ICA	Automated plaque stability prediction and plaque width assessment	Segmentation, Quantification, Classification	Sen/Spe vulnerable/stable plaque: 72.5/48.5; R thickness: 0.32
[38]	276/276	CUS	CCA, ICA, ECA	Plaque segmentation and effect of image standardization	Segmentation	DSC: 84.4 ± 8.1
[39]	491/512	CUS	CCA	Segmentation of IMT and quantification of SSI	Quantification	R SSI%: 0.928/0.704 internal/external test
[40]	134/659	CUS	na	Plaque segmentation using a U-Net integrating attention mechanism	Segmentation	DSC: 82.54 ± 0.73
[41]	204/407	CUS	CCA	Lumen segmentation and SSI measurement using DL with attention mechanisms	Classification, Quantification	R SSI%: 0.92/0.8 DL1/DL2; AUC stenosis risk: 0.88/0.98/1 and 0.93/0.97/1 low/moderate/high risk for DL1 and DL2, respectively
[42]	144/506, 497/636	CUS	na	Plaque segmentation using an image registration-based self-supervised learning method	Segmentation, Quantification	DSC: 80.25 ± 9.57 (n = 10), 85.40 ± 6.67 (n = 33), 86.72 ± 5.72 (n = 50), 89.18 ± 4.56 (n = 100); R TPA: 0.985
[43]	>200/na	CUS	na	Plaque segmentation	Segmentation	DSC: 93.81
[44]	157/5662, 8/4889	CUS	na	Real-time plaque segmentation using a Spatial–Temporal Feature Filter and multiscale features	Segmentation	DSC: 85.98 (DB1), 89.44 (DB2)
[45]	na/84 (baseline) + 84 (FU) volumes	3D US	na	Evaluation of 3D-US image segmentation workflow and VWV/WVT quantification	Quantification	R VWV: 0.69/0.77 patient-/time-based; R VWT: 0.69/0.73 patient-/time-based
[46]	144/510	CUS	na	Combination of CNN models to improve accuracy and segmentation performance	Segmentation, Quantification	DSC: 88.88 ± 4.36, R TPA: 0.967
[47]	117/117	CUS	na	Optimization of plaque segmentation using texture information from US images	Segmentation	Acc mean: 80.33
[48]	na/568	Doppler US	na	Plaque segmentation and vulnerable/stable classification with DL system	Segmentation, Classification	Acc and AUC vulnerable/stable: 92.94/89.41 and 0.915/0.853 Inception_v3/ResNet50
[49]	245/na	CUS + CEUS	na	Classification of fibrous cap integrity of plaque	Segmentation, Classification	Acc and AUC vulnerable/stable: 92.35 and 0.935
[50]	144/510, 497/638	CUS	na	Plaque segmentation and TPA quantification in limited labeled training dataset using self-supervised learning	Segmentation, Quantification	DSC/R TPA: 80.61 ± 9.75/0.852 DB1, 84.91 ± 6.75/0.936 DB2, 85.69 ± 6.71/0.957 DB3, on external Zhongnan test
[51]	2379/8448	CUS	CCA, Bulb	Framework for quantification of IMT and classification based on presence/absence plaque	Classification	Acc plaque yes/no: 97/81 CCA/bulb
[52]	99/970, 190/379, 50/300	CUS	ICA, CCA	Plaque segmentation and stroke risk assessment with attention-based DL model	Segmentation, Quantification	DSC: 89.90 ± 3.69 ICA, 86.50 ± 5.94 CCA; R TPA: 0.99 ICA, 0.96 CCA
[53]	99/970, 190/379, 50/300	CUS	ICA, CCA	Plaque segmentation and measurement of TPA by hybrid DL architectures	Segmentation, Quantification	DSC min–max: 78.88–88.37 (CCA); 75.14–90.02 (ICA)
[54]	na/4384, na/431	CUS	na	Encoder–decoder architecture for automated plaque segmentation	Segmentation	DSC: 83.65
[55]	90/115	CUS	ICA, CCA	Improvement of plaque segmentation and estimation of TPA using transfer learning	Segmentation, Quantification	DSC: 82.1 ± 5.3
[56]	108/67	CUS + Doppler US	CCA	Identification of plaque components using a patch-based DL method	Segmentation	JSC: 67.34/25.17/26.54 fibrous/lipid/calcified plaque
[57]	190/379	CUS	CCA	High-risk plaque segmentation and TPA quantification with a hybrid DL method	Segmentation, Quantification, Classification	DSC: 88.23 ± 7.75, R TPA: 0.82 (UNet), 0.85 (SegNet-UNet); AUC stenosis risk: 0.94 (UNet), 0.93 (SegNet-UNet)
[58]	na/2096	CUS	na	Intima–media segmentation and plaque thickness assessment based on 2D images	Quantification	R2 thickness: 0.982
[59]	99/970	CUS	left/right ICA	Comparison among solo DL and hybrid DL models for plaque segmentation and quantification	Segmentation, Quantification	DSC/R TPA: 88.98 ± 1.04/0.974 (cross entropy-loss), 86.98 ± 0.74/0.978 (DSC-loss)
[60]	165/630, 50/300	CUS	CCA	Investigation of “unseen AI” paradigm for plaque segmentation across ethnic groups	Segmentation, Quantification	DSC/R TPA: 78.38 ± 10.11/0.8 (UnseenAI-1), 82.49 ± 8.44/0.87 (UnseenAI-2), 86.89 ± 6.43/0.92 (SeenAI/Mixed)
[61]	295/25,289	CUS	na	Multi-plaque tracking and segmentation in US-video	Segmentation	DSC: 78 ±15 (MSTUnet), 69 ± 13 (Dual Attention U-Net), 83 ± 12 (Test1), 80 ± 2 (Test2)
[62]	144/510, 497/638	CUS	CCA, ICA, ECA	Plaque segmentation and TPA measurement	Segmentation, Quantification	R TPA: test1 0.989/0.987 (OP1/OP2), Zhongnan 0.915/0.942 (OP1/OP2)
[63]	144/510, 497/638	CUS	CCA, ICA, ECA	Automated plaque segmentation and TPA measurement	Segmentation, Quantification	DSC: 83.3 ± 10.0 (DB1), 85.3 ± 8.3 (DB2), 85.0 ± 7.8 (DB3); R TPA: 0.972 (Zhongnan)
[64]	204/250	CUS	CCA	Joint detection and measurement of VWT and PB using a two-stage model	Segmentation, Quantification	R TPA: 0.89 two-stage DL
[65]	101/862	CUS	ICA, CCA, bifurcation	Plaque segmentation in severely stenotic cases	Segmentation	DSC: 55 ± 19 dilated U-Net, 84 ± 5 semi-dilated U-Net
[66]	2379/8484, 27/4751	CUS	CCA, Bulb	Plaque detection and IMT measurement using single-step semantic segmentation	Classification	Acc plaque presence: 96/78 CCA/bulb
[67]	na/1007, 21/21	3D US	CCA, bifurcation	Segmentation of MA/LI borders from 3DUS for VWV measurements	Quantification	R VWV: 0.945 external SPARC dataset
[68]	38/144	3D US	CCA	Segmentation of MA/LI borders from 3DUS for VWV measurements	Quantification	R VWV: 0.96
[69]	204/407	CUS	CCA	SSI measurement and risk stratification in diabetic patients	Classification, Quantification	R SSI%: 0.93/0.94/0.93 DL1/DL2/DL3; AUC stenosis risk: 0.9/0.94/0.86 DL1/DL2/DL3

Abbreviations: US, Ultrasound; CUS, Conventional ultrasound; CEUS, Contrast-enhanced ultrasound; CCA, Common carotid artery; ICA, Internal carotid artery; ECA, External carotid artery; na: not available; RCCM-Net, Region and category confidence based multi-task network; SSI, Stenosis severity index; PB, Plaque burden; IMT, Intima–media thickness; DL, Deep learning; TPA, Total plaque area; VWV, Vessel wall volume; VWT, Vessel wall thickness; CNN, Convolutional neural network; AI, Artificial intelligence; MA, Media–adventitia; LI, Lumen–intima; DSC, Dice similarity coefficient; Acc, Accuracy; R, Correlation coefficient; Sen, Sensitivity; Spe, Specificity; AUC, Area under the curve; OP, Operator; DB, Database; JSC, Jaccard similarity score; ResNet, Residual network; SegNet, Semantic segmentation; MSTUnet, Multi-stream similarity learning network.

for carotid and

Table 2. Main characteristics of the 15 original articles included in the analysis and focused on coronary vessel IVUS imaging.

Ref.	N Subjects/N Images	US Imaging	Aim of the Study	Proposed Task	Main Results
[70]	11/5625, 5/791	IVUS	Plaque detection and classification using a toolbox for semi-automatic annotation and PB quantification	Quantification	R PB: 0.951
[71]	10/2175	IVUS	Plaque detection and classification using a DL hybrid technique	Segmentation, Classification	DSC: 96.16 ± 1.3 (Dice-loss), 98.88 ± 1.0 (Focal-loss), 98.97 ± 2.3 (Tversky-loss); Acc plaque/calcification: 97.33/96.94
[72]	1240/191,407	IVUS	PB quantification based on EEM/lumen segmentation	Quantification	ICC volume: 0.94
[73]	292/35,930	IVUS	Postprocessing pipeline for automated calculation of clinical parameters of vessel and plaque	Quantification	R PB: 0.862
[74]	153/68,549	IVUS	Vessel segmentation and PB quantification reducing scale-dependent interference	Quantification	R PB: 0.93
[75]	1063/na	IVUS	Quantification of vessel and plaque parameters from lumen/EEM segmentation	Quantification	R PB: 0.86/0.85 OP1/OP2
[76]	70/23,774, 77/435	NIRS—IVUS	EEM/lumen segmentation using POST-IVUS framework and TPA quantification	Quantification	Modified William index: 1.248
[77]	na/4197	IVUS	3D reconstruction and calcified/noncalcified plaques characterization tool	Segmentation, Classification	Characterization Acc: 91.43
[78]	100/5089	IVUS	Classification of vascular lesions including plaques (fibrous/lipid/calcified)	Segmentation, Classification	DSC and Acc: 86.29 ± 1.03/84.31 ± 0.99/84.48 ± 1.18 and 95.08/95.47/94.33 for fibrous/lipid/calcific
[79]	10/2175	IVUS	Classification of plaques and calcification	Classification	Acc plaque/calcification: 96.45/97.94
[80]	18/1746	IVUS	Lumen/MA segmentation using feature pyramid network and PB quantification	Segmentation, Quantification	R PB: 0.976
[81]	63/13,435	IVUS	Segmentation of lumen/vessel and PB estimation using ML approach	Quantification	R PB: 0.95
[82]	na/175 pullbacks	IVUS	Comparison among different CNNs for lumen/EEL segmentation and TPA quantification	Quantification	R TPA: 0.98
[83]	65/824,750	IVUS	Segmentation of lumen in real-time high-resolution IVUS images and TPA estimation	Quantification	R PB: 0.93
[84]	18/713	IVUS	Segmentation of MA/lumen/calcific plaque	Segmentation	DSC: 59 ± 39 (0–10% plaque), 74 ± 22 (0–10%), 84 ± 9 (>30%), 67 ± 15 (mean)

Abbreviations: US, Ultrasound; IVUS, Intravascular ultrasound; na: not available; PB, Plaque burden; EEM, External elastic membrane; DL, Deep learning; TPA, Total plaque area; CNN, Convolutional neural network; ML, Machine learning; MA, Media–adventitia; DSC, Dice similarity coefficient; Acc, Accuracy; R, Correlation coefficient; ICC, Intraclass correlation coefficient; OP, Operator.

for coronary), and depending on the DL network tasks, which are presented in sixth and fifth columns of Table 1 and Table 2, respectively. Moreover, the presentation of the quantitative results was organized according to task-specific metrics related to plaque analysis. In the case of multiple sets of evaluation results being reported by a paper, all relevant metrics were reported, with a preference for those providing external testing. Additionally, when manual segmentations from more than one operator were available, the results were reported individually rather than as average values. Furthermore, for studies reporting the same indexes, means and standard deviation values were assessed and graphed. Given the heterogeneity in imaging protocols, patient populations, and annotation methodologies across the various studies, a meta-analysis or a pooled statistical analysis of all the results was not feasible. This limitation highlights a critical challenge in the current landscape of medical image analysis: the lack of standardized benchmarks and datasets hinders the direct comparison and generalization of findings across different research groups. Therefore, our analysis focuses on qualitative synthesis and identifying trends and gaps in the existing literature, rather than providing a quantitative meta-analysis of effect sizes.

3. Results

3.1. Search Results

The systematic search identified a total of 402 publications across three databases: PubMed/Medline electronic database (64), Scopus (184), and WoS (154). After removing duplicates (N = 169), the screening was narrowed down to 233 publications. Among these, 90 documents were deemed ineligible for our purposes and excluded from the final screening, including abstracts, conference proceedings/reviews, reviews, and other types of publications that were not original papers. After reviewing the titles and abstracts, and the full text when necessary, a further 90 records were rejected. Ultimately, a total of 53 publications [32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84] met the inclusion criteria and were included in the systematic review.

Figure 1 shows a flow diagram that illustrates the systematic identification, screening, and inclusion of articles.

The main characteristics of the reviewed papers are summarized in Table 1 and Table 2, splitting studies based on the vascular district of interest—carotid and coronary, respectively.

Although the search strategy did not impose any date restrictions on publication, all selected papers were published within the last six years, specifically between 2019 and 2025 (Figure 2a). Figure 2b shows the number of studies categorized by the tasks related to plaque, which include segmentation, stratification or plaque severity classification, and the extraction of quantitative parameters from plaque.

The following subsections present results based on the two only anatomical areas of interest uncovered by our analysis, namely carotids and coronary vessels, along with the US modality primarily used in each case, IVUS and CUS, respectively. Out of the 53 selected studies, the majority (N = 38, 72%) focused on applying DL techniques to segment the carotid artery’s plaques and/or lumen in CUS images. The remaining 15 (28%) papers focused on coronary IVUS. Furthermore, the selected papers will be analyzed focusing on the DL tasks on plaques, i.e., segmentation, classification, and/or quantification/quantitative parameter extraction.

3.2. Carotid US

Among the 38 proposed deep learning (DL) segmentation techniques applied in US images, 27 (71%) were specifically focused on identifying atherosclerotic plaques [32,33,35,37,38,40,42,43,44,46,47,48,49,50,52,53,54,55,56,57,59,60,61,62,63,64,65].

Across these 27 studies, 14 papers not only concentrated on plaque segmentation but also derived the quantitative parameters of carotid arteries. For example, [37,42,46,50,52,53,55,57,59,60,62,63,64] quantified the total plaque area. Additionally, Jain et al. [57] assessed patients’ risk stratification depending on the plaque area.

Moreover, four studies [32,35,48,49] integrated DL-based techniques for both plaque segmentation and classification tasks. Papers [32] and [35] focused on classifying plaques into hypo-echoic, hyper-echoic, and mixed-echoic categories. Paper [49] focused on classifying fibrous cap integrity, while paper [48] differentiated between vulnerable and stable plaques.

Out of the 38 papers focused on carotid US, 10 studies (26%) utilized DL methods to segment vascular structures other than plaques. Among these, three studies [34,51,66] first applied DL-based vessel segmentation and then aimed to classify images for the presence or absence of plaques. Notably, Vila et al. [66] for the first time introduced a novel single-step approach using DenseNet for the semantic segmentation of carotid arteries, along with the detection of atherosclerotic plaques based on a large dataset of more than 8000 images. Additionally, the authors evaluated the generalization capability of their model by testing it on an external dataset (NEFRONA dataset), achieving plaque detection accuracy rates of 96.45% for the common carotid artery (CCA) and 78.09% for the bulb. These results were obtained using the same dataset referenced in study [51] but with a different DL architecture.

Two additional papers published by the same research group [41,69] proposed a classification approach to assess plaque risk levels by categorizing images into low, moderate, and high risk. Specifically, they quantified the severity stenosis index (SSI) and defined low risk as SSI < 25%, moderate risk as 25% ≤ SSI ≤ 50%, and high risk as SSI > 50%. In particular, in paper [41], the authors applied an attention mechanism-enhanced DL model to their previous implementation [64,69], achieving improved results.

Additionally, three papers used 3D-US imaging acquisition systems [45,67,68]. Notably, Zhou et al. [68] was the first group to propose a DL segmentation approach for plaque on 3D-US images. They introduced a semi-automatic CCA segmentation method using dynamic CNN (DCNN) and U-Net to segment the media–adventitia and lumen–intima boundaries in cases with carotid stenosis > 60% [68]. One year later, the same group proposed another DL segmentation network combining a 3D DCCN with a continuous max-flow module to segment the CCA from a manually identified region of interest (ROI) [67].

A few papers have used imaging techniques that differ from conventional B-mode US or have combined them with other methods. In particular, Doppler US was used in two studies [48,56], providing insights into blood flow and velocity, which can aid in assessing the presence of stenosis. A multimodal approach, integrating contrast-enhanced ultrasound (CEUS) and CUS, was proposed in study [49]. This approach demonstrated that enhanced contrast and visualization were useful for evaluating the carotid artery wall and lumen.

Interestingly, among the selected studies, seven (18%) incorporated attention mechanisms [33,35,40,41,49,52,61] to improve model performance for specific tasks. Attention mechanisms integrated with UNet [25] were first introduced by Oktay et al. [85] for pancreas segmentation. These attention mechanisms allowed the model to focus on the most relevant features of the input data and enhance the differentiation among various anatomical structures (such as plaque, lumen, and vessel) found in complex US images.

Numerous studies used attention-enhanced variants of U-Net. For instance, [33] integrated a Dual Pooling Self-Attention module specifically for plaque segmentation. Additionally, Xie et al. [40] introduced a Convolutional Attention Shrinkage Module (CASM) to capture long-range dependencies, while Biswas et al. [41] presented a MultiNet2.0, a U-Net that employs a VGG13 encoder and an attention mechanism to improve the segmentation of the lumen–intima border. Another notable development is WAL-Net, proposed in [35], which integrates an attention mechanism to enhance segmentation and provide auxiliary information for classification.

Building on these advancements, more recent studies have explored Dual Attention U-Net architectures, which enhance feature selection by jointly modeling spatial and channel-wise dependencies. Li et al. [61] were the first to introduce the Automatic Multi-Plaque Tracking and Segmentation (AMPTS) framework, which consists of three interconnected modules. The first module utilizes a Dual Attention U-Net to simultaneously detect multiple plaques and their surrounding vessels. This approach was also employed in another study by the same research group [49] for assessing the fibrous cap integrity, further demonstrating the effectiveness of Dual Attention mechanisms in vascular imaging.

3.3. Coronary Intravascular US

Intravascular ultrasound (IVUS) provides a direct visualization of the coronary vessel wall and is widely used to assess atherosclerotic plaque formation and the degree of stenosis. Among the 53 studies reviewed, 15 (28%) utilized deep learning techniques for IVUS image analysis. The primary goal of these studies was to delineate the boundaries between the lumen and the external elastic membrane (EEM), which is essential for accurate plaque detection. Proper identification and segmentation of the lumen and EEM borders are crucial steps in conventional IVUS interpretation and are prerequisites for the quantitative assessment of coronary artery disease. While only one study focused exclusively on plaque segmentation [84], several others incorporated plaque classification [71,77,78,79] or the assessment of quantitative parameters [70,72,73,74,76,81,82,83].

Some studies took a more advanced approach by not only segmenting plaques but also classifying different types of plaques. We identified four studies with this dual objective [71,77,78,79]. Prajapati et al. (2023, 2025) [71,79] classified plaques on their compositional differences, distinguishing between plaques and calcified samples. Similarly, Meng et al. [78] and Kyriakidis et al. [77] assigned clinical significance to plaque types by differentiating between calcified, fibrous, and lipid-rich plaques.

Ten studies concentrated on quantifying atherosclerotic plaque parameters, offering clinically relevant insights [70,72,73,74,75,76,80,81,82,83,84]. These studies not only segmented the lumen and EEM but also extracted the key quantitative metrics essential for assessing disease severity. Commonly measured parameters included the lumen cross-sectional area (CSA) and EEM CSA, which were used to calculate the plaque burden (PB)—the percentage of the EEM area occupied by atherosclerotic plaque—an important clinical indicator [70,73,74,75]. Other studies also computed the atherosclerotic plaque area [76,82,83], while others measured the minimum and maximum diameters of both the lumen and EEM, as well as the minimum and maximum plaque thickness. These comprehensive evaluations enabled volumetric calculations [72,75,80].

For instance, Kim et al. [72] validated the clinical relevance of model-derived measurements by correlating IVUS parameters with clinical outcomes. Their findings indicated that a plaque atheroma volume exceeding 52.5% was the best predictor of three-year cardiac mortality, while a plaque burden at the minimal lumen area greater than 76.5% was associated with revascularization of the culprit-related target vessel. Similarly, Zhu et al. [80] quantified 12 clinical parameters from the American College of Cardiology guidelines, and compared them with annotations from clinical experts, obtaining a strong correlation (R > 0.90) for at least 11 parameters. This confirmed the robustness of their model.

In these studies, segmentation accuracy is a crucial factor for the reliability of classification and quantitative assessments, highlighting the clinical relevance of the application of deep learning methods in IVUS-based CV diagnostics.

3.4. Quantification of Plaques

Focusing on the ability of various DL models to extract quantitative geometric parameters (e.g., plaque area, volume, maximum plaque thickness, plaque burden, etc.), here, we present their performances. The above-mentioned parameters are crucial for clinicians assessing risk and monitoring progression over time. The manual evaluation of these markers is often very time-consuming. In Figure 3a–c, we have categorized the most representative studies based on the quantitative parameters they extract, specifically total plaque area, SSI, and PB, respectively.

Figure 3a presents correlation values (R) between manual and DL-based measurements of total plaque area (TPA). TPA is widely recognized as a strong predictor of CV events because it reflects the overall burden of atherosclerosis, rather than just focal stenosis. Several models achieve correlation values exceeding 0.98, indicating high reliability in plaque quantification. Notably, Attention-UNet applied to ICA images achieves an R value of 0.99 [52], while both Inception-UNet (R = 0.99) and UNet (R = 0.99) on ICA, as reported in [53], also demonstrate excellent performance. The U-Net architecture shows values of R = 0.989 for OP1 and R = 0.987 for OP2 [62], outperforming most other architectures. A notable trend in the performance of ensemble models is evident, as illustrated in [63]. Four ensemble tests demonstrate strong correlation: UNet++ Ensemble—test1 (R = 0.985), UNet++ Ensemble—test2 (R = 0.985), UNet++ Ensemble—test3 (R = 0.988), and UNet++ Ensemble—test4 (R = 0.972).

Among studies focusing on the quantification of SSI% [34,39,41,50,69], most systems exhibited correlation coefficients higher than 0.9 when comparing predicted values to ground truth (manual assessments) (Figure 3b).

Lastly, some research has examined the quantification of plaque burden as defined in Section 3.3, with results shown in Figure 3c.

3.5. Plaque Characterization/Classification

In the context of atherosclerotic plaques, it is clinically valuable to classify these plaques based on factors such as their severity percentage, vulnerability/stability, or type of plaque composition (e.g., collagen, lipids, calcium depots). Indeed, US imaging can assist in discriminating the different cellular components due to their differing echogenicity, making tissue-based classification feasible. Among the papers reviewed, three studies [37,48,49] focused on classifying plaques into vulnerable and stable types.

Several studies proposed methods for distinguishing between images with the presence/absence of plaques. In [34] the authors proposed two different approaches, frame-based and video-based, for evaluating the performance of their algorithm. They developed a CNN designed for integration into a freehand portable US system. This system processed 2D images using a U-Net architecture and then reconstructed a 3D volume of the artery. Regarding the classification of different types of plaques based on echogenicity (hyper/hypo/mixed), studies [32,35] achieved remarkable results with both models demonstrating over 90% accuracy and over 85% precision. In terms of vulnerability characterization, Inception_v3 proposed in [48] demonstrated the highest accuracy of 93%, while ResNet50 [48] reached 89% and BP-Net [49] attained 92%. In comparison, the study reported in [37] had a lower accuracy of 53%. Lastly, the authors in [78] presented a classification method for plaque types using IVUS images. This study showed impressive results, with all classifiers exceeding 95% accuracy, 95% specificity, and a 98% F1 score across the test sets.

3.6. Direct DL Plaque Segmentation

Figure 4 shows the mean values of the Dice similarity score (DSC) from various studies focused on direct plaque segmentation. This refers to DL models that segment plaque directly from the input image in a single step and differs from multi-step pipelines, where the quantification and/or classification of plaque occurs after segmenting lumen vessels or other structures. DSC mean values range from 55% [65] to 98.96% [71], indicating a broad spectrum of performance among the methods evaluated. Notably, a significant portion of the studies, particularly the majority, cluster around the mean, especially within the 80–90% range, suggesting that most methods achieve relatively strong segmentation results. However, few studies report lower performance, with DSC values below 70% [61,65,84]. For instance, in reference [65], the authors reported a DSC of 55% for a fully automatic DL system, which improved to 84% for a semi-automatic approach where a rectangular region of interest (ROI) box was provided around the plaque. In reference [84], the authors found a DSC of 59% with a wide standard deviation of 39% in images with calcified proportions of less than 10%. Interestingly, they achieved a DSC of 84% when calcifications exceeded 30%. It is also important to note that the standard deviations vary significantly across different studies.

4. Discussion

The application of AI technologies to medical imaging has advanced rapidly in recent years, significantly impacting both clinical practice and research. The emergence of DL architectures, coupled with the development of more complex neural networks, has created new opportunities to overcome the main limitations of traditional methods for image analysis. This progress aims to achieve greater precision and time-efficiency than current options.

As the burden of stroke and CV disease continues to rise, imaging modalities for diagnosing and predicting CV conditions have become among the earliest research areas for applying DL technologies. Recent research has focused on the tools and information required by medical professionals to predict CV risk and outcomes in patients with atherosclerosis. This has led to the exploration of DL architectures in atherosclerosis imaging to assist in numerous tasks, such as automated segmentation, classification, and quantitative characterization of atherosclerotic lesions from images [86]. When comparing carotid B-mode ultrasound (CUS) and intravascular ultrasound (IVUS), there are some minor differences in image quality and the characterization of plaque morphology. CUS images often suffer from poor contrast, inconsistent brightness, and anatomical variability, making it challenging to delineate plaques clearly. In contrast, IVUS provides a direct visualization of the vascular structure but may also face challenges with inadequate contrast and brightness inconsistencies, especially in areas with calcifications or shadowing. Additionally, catheter movement or blood flow can introduce further artifacts in the images. Despite the distinct challenges posed by each imaging modality, specialized deep learning (DL) solutions have been developed to address these issues. Current DL methods effectively address similar challenges across both US modalities by facilitating automated segmentation and reducing operator dependence, while also enabling the extraction of quantitative data. However, these DL methods do share common limitations, including data scarcity and a lack of standardization between the two US techniques. Implementing accurate and efficient methods for these tasks would help in defining plaque-based risk factors more effectively and would inform tailored treatment strategies for patients with coronary/carotid atherosclerotic disease. As new technologies have emerged, DL-based methods have been introduced in US imaging to improve plaque segmentation and, more importantly, to provide a comprehensive characterization and quantification of atherosclerotic plaques parameters (e.g., area, volume, composition, and burden) that reflect the severity and progression of the disease over time. Furthermore, DL-based segmentation models have been proposed as part of advanced strategies for plaque classification, aiming to enhance plaque staging and prognostication with the goal of optimizing current care algorithms [24,87].

Due to the similarity of DL network tasks among the reviewed papers, the main performance metrics were collected, and the results were summarized by grouping studies based on related tasks concerning plaques. This includes the performance of segmentation, risk stratification, or severity classification, and/or quantitative parameter prediction. This approach was preferred as it allowed scientists to evaluate the effectiveness and robustness of their own proposed methods against the state-of-the-art.

In the context of plaque segmentation, as shown in Figure 4, the performance of DL segmentation algorithms offers valuable insights into their robustness and reliability. The DSC mean value of 84.40 indicates a generally high level of agreement between the segmentations predicted by DL and the actual ground truth, as defined by manual plaque segmentation. However, the variability observed in the results underscores the impact of factors such as small sample sizes, differing model architectures, and the absence of external tests. These factors can lead to lower generalization capabilities of the models. While the average DSCs suggest strong overall performance, the differences across various studies emphasize the need for standardized evaluation protocols and the use of more diverse datasets. This could help ensure the broader applicability and reliability of the models.

The task of plaque classification has been addressed in fifteen studies, as detailed in Section 3. Notably, studies that focused on classifying plaque composition [32,35,71,78,79] achieved accuracies exceeding 90%. Additionally, two studies [48,49] that aimed to classify plaque vulnerability also reported impressive results, with both achieving an accuracy of 92%. In contrast, research presented in [37] indicated a lower performance in differentiating between stable and progressive plaques, with an accuracy of only 53%.

This systematic analysis also highlighted semi- and fully automatic models recently developed to assist clinicians in extracting the quantitative parameters of plaque. The mean correlation value between the plaque area predicted by the algorithms and the manual quantification was found to be R = 0.948 (Figure 2a), indicating a strong agreement between DL-based quantification and expert assessments. The highest accuracy for total plaque area quantification was reported in [52], where the Attention U-Net was used on ICA images, further supporting the idea that attention mechanisms enhance model prediction performance. Conversely, lower performance was noted for total plaque area quantification using Autoencoder (with a correlation of R = 0.76) [53] and SegNet-UNet (R = 0.85) [57], suggesting the potential limitations of autoencoder-based approaches and simpler network architectures.

Research on attention mechanisms has focused on segmenting vascular structures and plaques, with several studies [33,35,40,41,49,52,61] highlighting the increasing use of attention-enhanced U-Net models [33,35,40,41,52] and a shift toward the Dual Attention U-Net architecture [49,61].

Overall, the comments above demonstrate that the application of DL methodologies in US image analysis for atherosclerotic plaques holds transformative potential for clinical practice. It can improve accuracy, reduce inter-operator variability, and enhance the prediction of risk through advanced plaque classification.

Another crucial factor in developing DL models is the availability of large datasets. Among the 53 studies reviewed, only a handful utilized datasets with an “adequate” number of patients to effectively train and test DL models (i.e., the more than 500 patients in [32,35,42,50,51,54,62,63,72,75]). Notably, six studies used data from the same dataset [32,35,42,50,62,63], while two others also shared datasets [51,66].

To address the challenge of low data availability, researchers often employ data augmentation strategies to increase the dataset size and reduce the risk of overfitting. This process involves applying a series of basic transformations to the original data, such as rotations, zooming/scaling, and movements along x- and y-axes. However, it is important to note that data augmentation does not equate to a true increase in new and independent data, and often neural networks may not benefit from an excessive amount of augmented data. Among the reviewed studies, 40 out of 53 (75%) reported utilizing data augmentation strategies [33,36,37,38,39,40,41,42,43,44,45,48,49,51,52,53,54,55,56,57,58,60,61,62,64,65,66,68,69,70,71,72,73,76,77,78,79,80,81,83]. Of these, 29 studies applied rotation, 21 used horizontal and/or vertical flips, 6 employed cropping, and 8 studies applied x/y translations.

Another crucial aspect of developing DL systems is the availability of external tests to assess a model’s generalization ability beyond the training environment. While internal tests often yield high performance, they may not fully capture the variability and complexity of real-world data. Conducting external validation using independent datasets helps identify potential overfitting and ensures that the model performs reliably under diverse conditions and with different input data. Notably, as reported in [39], the performance dropped significantly from R = 0.93 to R = 0.70 when moving from internal to external testing. These results highlight the importance of external tests in evaluating the “real” generalization capability of DL-based models. The gap between the theoretical performance achieved in controlled research settings and real-world clinical utility poses a significant challenge. This discrepancy arises from several factors: the inherent variability of clinical data, which can be influenced by different scanners, patient populations, and imaging protocols; the limited availability of diverse, multi-center datasets for training and validation; and the difficulty of adapting models developed on specific datasets to the broader clinical population.

One reason for the lack of external datasets to validate models may be the limited availability of large datasets, which are often collected from one or a few clinical centers and are commonly acquired using one or a limited number of echographic machines.

The scarcity of publicly accessible datasets poses a significant challenge in this field. Among the reviewed studies, three [71,76,79] utilized a public and downloadable dataset (the IVUS dataset referenced in [88]). However, this dataset is relatively small, comprising two IVUS coronary datasets A and B, which include 77 and 435 images extracted from pullbacks of 22 and 10 patients, respectively. Large datasets are preferred to enhance the generalization capabilities of DL models.

In addition to dataset availability, access to the code and the presence of public repositories (e.g., https://github.com/ accessed on 15 March 2025) that include instructions for use are crucial. This access would allow researchers to test models on their own dataset or adapt and extend them for slightly different applications or tasks. In the present review, the code was publicly available in only seven studies [35,44,51,66,67,70,74], limiting the adoption of these models and hindering opportunities for external validation and further development by the AI research community. Despite the promising results, several critical limitations hinder the widespread adoption of these methods in clinical practice. One major issue is the scarcity of large, diverse, and well-annotated datasets. This shortage leads to models that may perform well on specific datasets but struggle to generalize across different patient populations and imaging devices. The problem is exacerbated by the absence of standardized protocols for data acquisition and annotation, which hampers the development of robust and universally applicable models. Additionally, the black-box nature of many DL models raises concerns among clinicians. They require interpretability and explainability to trust the diagnostic outputs and effectively integrate these findings into their decision-making processes. Finally, the significant computational demands for training and deploying these models can pose a barrier, especially for institutions with limited resources and infrastructure.

Advances in the clinical applications of DL architectures would benefit from the availability of open datasets and codes, as this would enhance the reproducibility and transparency of research methodologies and results. Although sharing such information could lead to significant progress in the field, it raises ethical and privacy concerns, particularly regarding sensitive healthcare data.

In addition to the ongoing optimization of DL-based models, it is important to note the technological evolution of US imaging, transitioning from conventional 2D B-mode to the increasing use and advancements of 3D-US imaging, as well as the multimodal approach that combines morphological (B-mode) and functional data (Doppler and Contrast mode). In this regard, DL architectural innovations could inspire new methodologies and hybrid approaches. However, only a limited number of studies [45,67,68] have incorporated 3D-US techniques into their research on carotid artery analysis. This underutilization may be attributed to several factors. First, 3D-US requires specialized equipment and expertise that may not be as readily available in many clinical settings as conventional 2D US. Additionally, the processing and interpretation of 3D images involve more complex computational methods, which can pose technical challenges for integration into DL models. Another contributing factor is the lack of large, publicly available datasets specifically designed for 3D-US carotid imaging, which limits the development and validation of robust AI-driven approaches. Future research should focus on strategies to overcome these barriers, such as increasing data availability, improving automated segmentation algorithms, and enhancing the clinical usability of 3D-US technology.

Limitation of the Review

One important limitation, as above mentioned, that affects most of the papers analyzed in the review is the limited sample size of the cohorts studied in the papers, and the lack of effective external tests of the models across the reviewed studies.

It is worth mentioning that the standardization of quantitative indicators is a critical aspect when evaluating the effectiveness of proposed methods. Since the criteria used to evaluate performance among reviewed articles were quite heterogeneous and varied, we did not perform any statistical test to support the results presented in this review.

Furthermore, the main limitation of this review is the lack of formal risk of bias assessment due to the heterogeneity of the included studies, and this may limit the overall quality of the available evidence.

5. Conclusions

In conclusion, the application of deep learning (DL) to the imaging analysis of plaque using US has progressed rapidly. This advancement enables the automated segmentation, classification, and quantification of atherosclerotic lesions, promoting early diagnosis and timely intervention. Although DL models have demonstrated promising performance, variability in results underscores the need for larger datasets, improved data accessibility, open-source codes, and external validation. These elements are essential for ensuring generalizability and reproducibility, as well as supporting further research advancements. Additionally, the use of advanced 3D-US remains limited. Future efforts should prioritize data sharing, methodological standardization, and the enhanced integration of advanced imaging modalities to improve clinical applicability.