AI-Enhanced Super-Resolution Handheld Ultrasound for Carotid Plaque Detection in Community Screening

Abstract

PURPOSE

Ischemic stroke is a leading cause of disability and death, and early identification of carotid atherosclerotic plaques is critical for stroke prevention in primary care and community settings. We examined whether artificial intelligence (AI)-enhanced handheld ultrasound (HHUS) could improve carotid plaque detection and stenosis assessment in community-based primary care screening.

METHODS

We developed a super-resolution reconstruction model (Hyper-CycleGAN) and validated it in a clinical cohort of 127 patients with 198 plaques. The validated model was then applied in a community screening setting with 117 participants and 153 plaques, using portable ultrasound as the reference standard. We evaluated agreement with the reference standard (Bland-Altman, intraclass correlation coefficient) and community diagnostic performance (detection rates, quadratic-weighted Cohen κ, confusion matrices, and standard metrics).

RESULTS

In community screening, AI-enhanced HHUS identified 94.8% of carotid plaques compared with 87.6% using standard HHUS. Agreement in stenosis grading with the reference standard improved, with weighted κ increasing from 0.492 to 0.836. Sensitivity for identifying ultrasound-defined vulnerable plaques increased from 47.4% to 63.2%, and specificity remained high.

CONCLUSIONS

AI-enhanced HHUS improved carotid plaque detection and stenosis assessment in community screening, supporting more accurate risk stratification and referral decisions in primary care. This approach could expand access to low-cost stroke prevention strategies in underserved communities.

Abstract available in: يبرع (Arabic); 中文 (Chinese); Francais (French); Deutsch (German); हिन्दी (Hindi); Indonesian (Indonesian); 日本語 (Japanese); Portugues (Portuguese); Español (Spanish)

INTRODUCTION

Ischemic stroke remains a leading cause of disability and death worldwide and poses a substantial burden on primary care systems responsible for long-term risk management and prevention.¹ Carotid atherosclerotic plaque is a key modifiable risk factor for ischemic stroke, particularly when plaque vulnerability or clinically significant stenosis is present.^2-4 Identifying individuals at increased risk before the onset of symptoms is therefore central to effective stroke prevention in community and primary care settings. Despite this need, carotid screening is not routinely implemented in primary care. Hospital-based ultrasound services are resource intensive and difficult to scale, and widespread referral of asymptomatic or high-risk patients could overwhelm specialty services, without clear benefit.^5,6 As a result, opportunities for early risk stratification and targeted prevention are often missed. Community health centers and primary care clinics represent an important setting for early identification of carotid disease, provided that screening tools are accessible, efficient, and sufficiently reliable to inform clinical decisions and referrals.^7-9

Carotid ultrasound is well established for plaque detection and stenosis assessment and is favored for its safety and low cost.^10,11 However, conventional ultrasound systems are often unavailable in community clinics because of cost, infrastructure, and maintenance requirements.¹² Handheld ultrasound (HHUS) devices offer a pragmatic alternative and are increasingly used in primary care, rural medicine, and community screening programs.^13-16 Their portability and affordability make them well suited for frontline use, but limitations in image quality could decrease confidence in plaque characterization and lead to underestimation of risk, particularly for subtle or small lesions.¹⁴

Recent advances in artificial intelligence (AI) have created opportunities to enhance the clinical utility of existing diagnostic tools without changing established workflows.^17-20 Super-resolution (SR) reconstruction methods can improve image clarity and structural detail in low-quality ultrasound images,^21,22 potentially strengthening the reliability of HHUS in primary care. Whether such approaches can meaningfully improve carotid plaque detection and stenosis assessment in real-world community screening remains uncertain.

In this study, we evaluated an AI-enhanced SR-reconstruction approach applied to HHUS images for carotid plaque screening in a community setting. We examined whether this approach could improve plaque detection, agreement in stenosis grading, and identification of ultrasound-defined vulnerable plaques compared with standard HHUS, with the goal of supporting more accurate risk stratification and referral decisions in primary care.

METHODS

Data Acquisition

We obtained ultrasound data from both retrospective hospital archives and prospective community screening. During the model development phase, we collected carotid ultrasound images from 3 tertiary hospitals in Hunan Province, China, from November 2023 to April 2024. This dataset included 500 paired handheld and conventional ultrasound image sets acquired from the same anatomic carotid segments during a single examination session, as well as 6,000 additional high-resolution conventional ultrasound images used for model training and internal validation. A separate hospital cohort of 127 patients with 198 carotid plaques was used for independent model validation before community deployment. For the community screening cohort, the target population consisted of adults aged ≥40 years who were able to complete the ultrasound examination and questionnaire independently. Participants were included if they had complete demographic and imaging data with adequate visualization of the carotid artery. Individuals who declined to provide informed consent, were unable to cooperate during the examination, or had incomplete or poor-quality images were excluded from analysis.

We determined stroke risk status with a community stroke risk screening form. Participants with a history of stroke or transient ischemic attack were classified as high risk. Participants without prior stroke or transient ischemic attack were classified as high risk if they had ≥3 of the following risk factors: smoking, hyperlipidemia, diabetes, physical inactivity, obesity (defined as body mass index ≥26 kg/m²), family history of stroke, hypertension, or atrial fibrillation or valvular heart disease. Participants with <3 risk factors but with hypertension, diabetes, atrial fibrillation, or valvular heart disease were classified as intermediate risk; the remaining participants were classified as low risk.

Conventional and handheld carotid ultrasound examinations in the hospital setting were performed during the same examination session, with handheld imaging obtained from corresponding carotid segments. In the community setting, HHUS served as the index imaging modality, and portable ultrasound was used as the reference standard. All examinations were performed by 2 vascular sonographers with >5 years of carotid ultrasound experience and formal vascular imaging certification. Before study initiation, sonographers were trained using a standardized carotid scanning protocol.²³ Participants were examined in the supine position with the neck mildly extended and the head turned contralaterally. Transverse and longitudinal images were obtained from the common carotid artery through the carotid bifurcation and proximal internal carotid artery. Probe position, depth, gain, and focal zone were optimized for each participant and kept as consistent as possible across HHUS and reference imaging. Images with inadequate visualization of the carotid wall or plaque boundary were excluded before analysis, and a senior sonographer reviewed image quality for protocol adherence. Handheld images were subsequently processed with the established SR model to generate the super-resolution handheld ultrasound (SR-HHUS) dataset. Conventional ultrasound in the hospital cohort and portable ultrasound in the community cohort were used as the best available reference standards. Reference images in the community cohort were interpreted independently by senior sonographers who were blinded to HHUS and SR-HHUS findings, and reference assessments were completed before comparative analysis to minimize incorporation bias. We performed HHUS with the SonoCon D6CL system, and the community reference standard used a SonoScape S9 Pro portable ultrasound with a higher-channel linear transducer and superior image resolution (Supplemental Appendix 1).

Ethics approval was obtained from the Institutional Review Board of Changsha Central Hospital, and all procedures adhered to the Declaration of Helsinki. Figure 1 provides an overview of the study workflow and data sources. For model development, paired handheld and conventional ultrasound images were acquired during the same examination session, with probes positioned sequentially on the same anatomic segments of the carotid artery under standardized scanning planes. Image pairs were matched by anatomic landmarks to minimize spatial mismatch. Community screening images were collected prospectively and were not included in model training or internal validation, ensuring complete separation between training and evaluation datasets.

Figure 1.

Illustration of Participant Inclusion, Image Acquisition, and Diagnostic Evaluation Flow

Conv = convolution; ECST = European Carotid Surgery Trial; G_AB = generator transforming domain A (HHUS) to domain B; G_BA = generator reconstructing domain A from domain B; HHUS = handheld ultrasound; HyperConv = hyperparameter-tunable convolution; L2 =mean squared error loss; SR = super-resolution; SR-HHUS = super-resolution handheld ultrasound; US = ultrasound.

Note: The proposed SR reconstruction model was first trained using retrospective ultrasound datasets, followed by internal validation on hospital-based imaging data and deployment in community screening programs.

Super-Resolution Model Development

We developed a controllable hyper cycle-consistent adversarial network (Hyper-CycleGAN) SR model as an image postprocessing tool to improve the quality of HHUS carotid plaque images. The model was not designed to independently diagnose carotid plaque or determine biological plaque vulnerability. Rather, it was intended to transform a low-resolution HHUS brightness mode (B-mode) image into a higher-resolution image with improved visualization of the plaque–lumen boundary, intima–media interface, and intraplaque echogenicity. The model input was a static 2-dimensional HHUS carotid image, and the output was a reconstructed SR-HHUS image of the same anatomic view.

The model used a bidirectional CycleGAN framework in which HHUS images were defined as the low-resolution domain and conventional ultrasound images were defined as the high-resolution reference domain. The architecture included paired generators and discriminators to learn image translation between the 2 domains while preserving the underlying carotid anatomy. We incorporated a HyperConv-based decoder (a hyperparameter-tunable convolutional module) with a controllable blur parameter to adjust the degree of SR reconstruction. During training, the model used adversarial, cycle-consistency, pixel-level, and perceptual losses to enhance image sharpness and decrease noise while minimizing distortion of plaque morphology. To improve robustness, high-resolution conventional ultrasound images were also degraded using simulated blur, grid, mosaic, and directional artifacts to generate low-resolution training examples. The generator backbone used 9 residual network residual blocks. Model training and offline inference were performed on a workstation running Ubuntu 22.04 (Jammy Jellyfish) and PyTorch 1.13.1, with an Intel i7-12700KF central processing unit, 64 GB RAM (Intel Corp), and 2 NVIDIA GeForce RTX 3090Ti graphics processing unit (NVIDIA Corp). In this study, SR reconstruction was applied to exported static HHUS B-mode images as an offline postprocessing step. Real-time inference on handheld ultrasound devices was not evaluated. Additional details on preprocessing, augmentation, architecture, training parameters, and software implementation are provided in Supplemental Appendix 1.

Image and Diagnostic Evaluation

The index test was SR-HHUS, and conventional vascular sonographers independently measured, for each plaque, the maximum longitudinal and transverse diameters, the maximal lumen diameter, and the residual lumen diameter at the site of stenosis on HHUS, SR-HHUS, and reference standard images. All image measurements and classifications were performed independently and blinded to the results of other imaging modalities. The stenosis rate was calculated according to the European Carotid Surgery Trial method as follows:

$$SR=(1-\frac{D_{stenosis}}{D_{theoretical}})\times 100\%$$

The residual lumen diameter was defined as the minimum lumen diameter at the site of stenosis, and the estimated original lumen diameter was defined as the theoretical normal lumen diameter at the same carotid segment before plaque encroachment. Stenosis severity is categorized consistently throughout the manuscript as mild (<50%), moderate (≥50% to <70%), and severe (≥70%).²⁴

Ultrasound-defined vulnerable plaque features were classified using B-mode echogenicity and morphologic characteristics. Plaques were considered to have ultrasound-defined vulnerable features if they were predominantly hypoechoic, had an irregular or ulcerated surface, or showed discontinuity of the plaque–lumen interface. Plaque echogenicity was categorized using the following 5-type scheme: type I, uniformly hypoechoic; type II, predominantly hypoechoic with small hyperechoic areas; type III, predominantly hyperechoic with small hypoechoic areas; type IV, uniformly hyperechoic; and type V, heavily calcified with acoustic shadowing.²⁵ Types I and II were classified as having ultrasound-defined vulnerable features, whereas types III to V were classified as stable-appearing plaques. The Carotid Plaque-Reporting and Data System was not applied. Because histology, computed-tomography angiography, and magnetic resonance imaging were not used, vulnerable plaque findings in this study refer to ultrasound-defined surrogate features rather than comprehensive biological plaque instability.

Statistical Analysis

The primary outcome of this study was the diagnostic performance of SR-HHUS in community screening, assessed by plaque detection performance and agreement in stenosis grading, with portable ultrasound used as the reference standard. Diagnostic performance for identifying ultrasound-defined vulnerable plaque features was analyzed as an additional outcome. Continuous variables are presented as mean (SD). Agreement between HHUS or SR-HHUS and reference ultrasound for plaque size measurements was evaluated using Bland-Altman analysis, with mean bias and 95% limits of agreement reported, and intraclass correlation coefficients (ICCs) used as a complementary measure of reproducibility. Proportional bias was assessed by regressing measurement differences on their mean values. Analyses were performed separately for hospital-based and community cohorts.

In the community cohort, plaque detection rates between HHUS and SR-HHUS were compared descriptively. Agreement in stenosis grading relative to the reference standard was evaluated using weighted Cohen κ statistics and confusion matrices. Diagnostic performance for vulnerable plaque features was assessed using sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve. This study was designed as a pragmatic diagnostic accuracy study conducted in a real-world community screening setting and is reported in accordance with the Standards for Reporting of Diagnostic Accuracy Studies 2015 guidelines.²⁶ No formal a priori sample size calculation was performed because the study was designed as a pragmatic community-based diagnostic accuracy study. However, a post hoc precision analysis was conducted at the plaque level. Assuming an expected sensitivity of 0.80 and a desired 95% CI width of ±10%, ≥62 plaques with the target condition would be required to estimate sensitivity with acceptable precision. The community cohort included 153 plaques, exceeding this threshold and providing adequate precision for estimating diagnostic performance metrics. All analyses were performed using IBM SPSS Statistics version 27.0. Participants with incomplete demographic or imaging data and those with poor-quality images were excluded prior to analysis, consistent with the predefined eligibility criteria. No missing index test or reference standard data were present in the final analytic dataset, and no data imputation was performed. Because some participants contributed >1 plaque, analyses were conducted at the plaque level without adjustment for within-participant clustering. Because the number of plaques per participant was small, clustering effects were considered unlikely to materially alter the primary findings.

RESULTS

Patient Characteristics

In the source community screening population, 450 participants aged ≥40 years underwent carotid ultrasound screening. Carotid plaque was identified by portable ultrasound in 170 participants, corresponding to a participant-level carotid plaque prevalence of 37.8% (Figure 1). After applying image-completeness and quality criteria, 117 participants with 153 analyzable carotid plaques were included in the final diagnostic accuracy analysis. The mean (SD) age was 70.4 (9.4) years, and 65 participants (55.6%) were female (Table 1). Hypertension and hyperlipidemia were present in 62.4% and 46.2% of participants, respectively; 30.8% had diabetes, and 18.8% reported current or previous smoking. Obesity, defined as BMI ≥26 kg/m², was observed in 17.1% of participants, and 20.5% reported a family history of stroke. Overall, 40.2% of participants were classified as being at high risk of stroke. According to the portable ultrasound reference standard, 144 plaques (94.1%) were classified as mild stenosis (<50%), 7 (4.6%) as moderate stenosis (≥50% to <70%), and 2 (1.3%) as severe stenosis (≥70%). In the hospital validation cohort (127 patients, 198 plaques), SR-HHUS showed improved agreement with conventional ultrasound for plaque diameter measurements.

Table 1.

Baseline Characteristics of Participants in Community Screening Cohort (n = 117)

Variable	Value
Age, y, mean (SD)	70.4 (9.4)
Female, No. (%)	65 (55.6)
Diabetes, No. (%)	36 (30.8)
Hypertension, No. (%)	73 (62.4)
Hyperlipidemia, No. (%)	54 (46.2)
Atrial fibrillation or valvular heart disease, No. (%)	5 (4.3)
Smoking history, No. (%)	22 (18.8)
Obesity (BMI ≥26 kg/m2), No. (%)	20 (17.1)
Family history of stroke, No. (%)	24 (20.5)
High risk of stroke, No. (%)	47 (40.2)

BMI = body mass index.

Notes: High risk of stroke was defined according to the community stroke risk screening criteria described in the Methods section. The source community screening population included 450 participants; 170 had carotid plaque on portable ultrasound, corresponding to a participant-level plaque prevalence of 37.8%.

Primary Diagnostic Performance in Community Screening

Plaque Detection Performance

In the community screening cohort, plaque detection performance of HHUS and SR-HHUS was evaluated using portable ultrasound as the reference standard. A total of 153 carotid plaques were identified on reference imaging. HHUS detected 134 plaques, corresponding to a detection rate of 87.6%, whereas SR-HHUS detected 145 plaques, yielding a greater detection rate of 94.8%.

All plaques identified on HHUS images were also visualized on SR-HHUS images. In addition, the 11 additional plaques detected after SR reconstruction were mainly small or low-contrast plaques with mild stenosis on reference imaging. These lesions were difficult to delineate on standard HHUS because of indistinct plaque–lumen interfaces or weak echogenic contrast.

Agreement in Stenosis Grading

Agreement in carotid stenosis grading between HHUS or SR-HHUS and the portable ultrasound reference standard was evaluated in the community screening cohort using quadratic-weighted Cohen κ statistics. Whereas HHUS showed moderate agreement with the reference standard (weighted κ = 0.492; 95% CI, 0.268-0.716; P < .001), SR-HHUS showed greater agreement in stenosis grading, achieving a weighted κ of 0.836 (95% CI, 0.672-0.999; P < .001), corresponding to excellent agreement (Table 2). Confusion matrix analysis further illustrated the distribution of stenosis classifications (Figure 2). Compared with HHUS, SR-HHUS was associated with fewer cases of stenosis underestimation. Using SR-HHUS, 97.4% of cases were classified concordantly with the reference standard, with 2.6% of cases differing by 1 grade and no 2-grade misclassifications observed.

Table 2.

Weighted Cohen κ Analysis of Agreement in ECST-Based Carotid Stenosis Grading

Method	Weighted κ	95% CI	P	Agreement level
HHUS vs portable US	0.492	0.268-0.716	P < .001	Moderate
SR-HHUS vs portable US	0.836	0.672-0.999	P < .001	Excellent

ECST = European Carotid Surgery Trial; HHUS = handheld ultrasound; SR-HHUS = super-resolution handheld ultrasound; US = ultrasound.

Note: Stenosis severity was categorized as mild (<50%), moderate (≥50% to <70%), or severe (≥70%) according to ECST criteria.

Figure 2.

Confusion Matrices Comparing ECST-Based Carotid Stenosis Grading Between HHUS and SR-HHUS

ECST = European Carotid Surgery Trial; HHUS = handheld ultrasound; SR-HHUS = super-resolution handheld ultrasound.

Note: Stenosis severity was categorized as mild (<50%), moderate (≥50% to <70%), or severe (≥70%). Each cell represents the number of cases and the corresponding row percentage (Row %), defined as the proportion within each reference-standard stenosis grade. Data for SR-HHUS shows a greater proportion of correctly classified cases along the diagonal line, indicating improved grading consistency and fewer underestimations of stenosis severity compared with HHUS.

Diagnostic Performance for Ultrasound-Defined Vulnerable Plaques

In the community screening cohort, diagnostic performance of HHUS and SR-HHUS for identifying ultrasound-defined vulnerable plaque features was evaluated using portable ultrasound as the reference standard. Both HHUS and SR-HHUS showed areas under the receiver operating characteristic curve (AUCs) significantly greater than 0.5 against the reference standard (Table 3). The AUC was numerically greater for SR-HHUS than for HHUS (0.794; 95% CI, 0.703-0.885 vs 0.724; 95% CI, 0.624-0.824), although the paired AUC comparison narrowly missed statistical significance by the DeLong test (P = .051). The SR-HHUS measurements showed greater sensitivity than HHUS (63.2% vs 47.4%, respectively; McNemar P = .021), while maintaining high specificity (95.7% vs 97.4%). Overall accuracy was 87.6% for SR-HHUS and 85.0% for HHUS.

Table 3.

Diagnostic Performance of HHUS and SR-HHUS for Identifying Ultrasound-Defined Vulnerable Plaque Features in Community Screening

Metric	HHUS	SR-HHUS	Comparison P value
AUC	0.724	0.794	DeLong P = .051
	(95% CI, 0.624-0.824)	(95% CI, 0.703-0.885)
Sensitivity	47.4%	63.2%	McNemar P = .021
Specificity	97.4%	95.7%	NA
Accuracy	85.0%	87.6%	NA

AUC = area under the curve; HHUS = handheld ultrasound; NA =not computed; SR-HHUS = super-resolution handheld ultrasound.

Note: Portable ultrasound was used as the reference standard. Both HHUS and SR-HHUS had AUCs significantly greater than 0.5 when evaluated against the reference standard (both P < .001). The DeLong test was used for the paired comparison of AUCs between HHUS and SR-HHUS.

Participant-Level Diagnostic Performance

We performed participant-level analyses by defining a positive result as detection of ≥1 carotid plaque, or ≥1 ultrasound-defined vulnerable plaque feature, in a participant. For any carotid plaque detection, HHUS detected ≥1 plaque in 103 of 117 participants, whereas SR-HHUS detected ≥1 plaque in 113 of 117 participants, corresponding to participant-level sensitivities of 88.0% and 96.6%, respectively (Supplemental Table 1). Specificity for any plaque detection could not be estimated because all participants in the final analytic cohort had reference-confirmed carotid plaque. For ultrasound-defined vulnerable plaque features, SR-HHUS showed greater participant-level sensitivity than HHUS (63.9% vs 50.0%) and slightly greater accuracy (84.6% vs 82.1%), with slightly lower specificity (93.8% vs 96.3%).

Agreement With the Reference Standard

Bland-Altman Analysis of Plaque Measurements

Bland-Altman analysis showed improved agreement between SR-HHUS and the reference ultrasound for plaque size measurements (Figure 3, Supplemental Table 2). For plaque length, the mean bias decreased from 0.99 mm with HHUS to −0.01 mm with SR-HHUS, and the 95% limits of agreement narrowed from −6.82 to 8.79 mm to −1.28 to 1.25 mm. For plaque thickness, the mean bias changed from −0.04 mm with HHUS to −0.02 mm with SR-HHUS, with corresponding 95% limits of agreement narrowing from −1.07 to 0.99 mm to −0.44 to 0.40 mm.

Figure 3.

Agreement of SR-HHUS and HHUS Measurements of Plaque Size With Conventional Ultrasound

HHUS = handheld ultrasound; SR-HHUS = super-resolution handheld ultrasound.

Note: Red dashed line denotes mean bias; blue dashed lines denote 95% limits of agreement.

Intraclass Correlation of Plaque Measurements

Intraclass correlation coefficients were used to assess the reproducibility of plaque size measurements obtained from HHUS and SR-HHUS relative to reference ultrasound (Table 4). In the hospital validation cohort, SR-HHUS showed greater reproducibility than HHUS for both long-axis plaque diameter (ICC, 0.993; 95% CI, 0.991-0.995 vs ICC, 0.749; 95% CI, 0.673-0.808) and short-axis plaque diameter (ICC, 0.967; 95% CI, 0.957-0.975 vs ICC, 0.810; 95% CI, 0.756-0.853). Similar findings were observed in the community cohort, where SR-HHUS showed greater ICCs than HHUS for long-axis diameter (0.971; 95% CI, 0.960-0.979 vs 0.867; 95% CI, 0.813-0.905) and short-axis diameter (0.956; 95% CI, 0.939-0.969 vs 0.844; 95% CI, 0.787-0.886).

Table 4.

ICC Analysis of Carotid Plaque Diameter Measurements for HHUS, SR-HHUS, and Reference US

Study phase	Comparison	Plaque long-axis diameter	Plaque short-axis diameter
Hospital validation	HHUS vs US	0.749 (95% CI, 0.673-0.808; P < .001)	0.810 (95% CI, 0.756-0.853; P < .001)
	SR-HHUS vs US	0.993 (95% CI, 0.991-0.995; P< .001)	0.967 (95% CI, 0.957-0.975; P < .001)
Community screening	HHUS vs US	0.867 (95% CI, 0.813-0.905; P < .001)	0.844 (95% CI, 0.787-0.886; P < .001)
	SR-HHUS vs US	0.971 (95% CI, 0.960-0.979; P < .001)	0.956 (95% CI, 0.939-0.969; P < .001)

HHUS = handheld ultrasound; ICC = intraclass correlation coefficient; SR-HHUS = super-resolution handheld ultrasound; US = ultrasound.

Secondary Analyses

Qualitative Assessment of Image Quality

Qualitative assessment of image quality showed that SR reconstruction improved the visual delineation of carotid plaques on HHUS images. Compared with original HHUS images, SR-HHUS images showed clearer plaque–lumen interfaces and more distinct visualization of the vessel wall structure, particularly at plaque margins (Figure 4). These improvements were most apparent in images with low contrast or indistinct boundaries on original handheld scans, where plaque morphology was difficult to interpret. Enhanced image clarity on SR-HHUS was consistently observed across representative cases, supporting the improved diagnostic performance and measurement consistency reported above.

Figure 4.

Comparison of Carotid Plaque Ultrasound Images and Model Reconstruction Images

HHUS = handheld ultrasound; SR-HHUS = super-resolution handheld ultrasound; US = ultrasound.

Note: A. HHUS shows a hypoechoic plaque with a blurred lumen–wall interface (arrow); B. SR-HHUS decreases speckle and sharpens plaque margins and the intima–media complex (arrow); C. Conventional US depicts clear plaque contours and lumen narrowing (arrow).

Measurement Bias and Proportional Bias

Assessment of proportional bias was performed by regressing the differences in plaque size measurements between HHUS modalities and the reference standard on their mean values. No clinically meaningful proportional bias was observed for SR-HHUS measurements, indicating that measurement differences did not systematically increase with plaque size. In contrast, HHUS measurements showed greater variability across the range of plaque sizes. These findings suggest that SR reconstruction decreased size-dependent measurement inconsistency, supporting the improved agreement and reproducibility observed in the preceding analyses. No adverse events related to the handheld or portable ultrasound examinations were observed in either the hospital validation cohort or the community screening cohort.

DISCUSSION

Results of this study show that AI-enhanced SR-HHUS can improve carotid plaque detection and stenosis assessment in a community screening setting, with clear implications for primary care. By enhancing image quality without altering scanning workflows, SR-HHUS increased plaque detection, improved agreement in stenosis grading, and improved sensitivity for identifying ultrasound-defined vulnerable plaque features compared with standard HHUS, while maintaining high specificity. The gain in plaque detection appeared most relevant for small or low-contrast plaques, most of which showed mild stenosis on reference imaging. These plaques were difficult to delineate on standard HHUS because of indistinct plaque–lumen interfaces or weak echogenic contrast. In community screening, improved visualization of these early or subtle plaques could support more complete vascular risk recognition, lifestyle counseling, medical risk-factor management, and referral for confirmatory vascular ultrasound when clinically indicated.

Carotid screening in asymptomatic general populations remains controversial, and our findings should not be interpreted as supporting indiscriminate population-wide screening.²⁷ Rather, the potential value of SR-HHUS lies in selective community-based risk assessment among older adults or individuals with vascular risk factors, where improved visualization might help identify patients who require confirmatory imaging, medical risk-factor optimization, or referral.^28,29 This distinction is important because SR-HHUS is intended to support triage and prevention in primary care, not to replace guideline-based clinical decision making.

Handheld ultrasound is increasingly used in primary care and community settings because of its portability and affordability, but concerns regarding image quality have limited its role in vascular risk assessment. The clinical contribution of SR-HHUS should be interpreted as image-quality enhancement rather than replacement of conventional vascular ultrasound. Unlike conventional HHUS, which is constrained by probe size, channel number, and onboard processing capacity, SR-HHUS applies postacquisition reconstruction to improve plaque–lumen boundary definition, vessel-wall continuity, and intraplaque echogenic contrast. Unlike simple interpolation or denoising filters, the Hyper-CycleGAN framework learns a mapping between low-quality handheld images and high-quality conventional ultrasound images while using cycle consistency to preserve the original anatomic structure. Our findings suggest that AI-based SR reconstruction can mitigate these limitations, strengthening the clinical utility of HHUS in settings where access to conventional imaging is constrained.^30,31 Improved image interpretability at the community level might help clinicians distinguish patients who can continue routine risk-factor management from those who need confirmatory vascular ultrasound or closer follow-up, thereby supporting more targeted use of specialty imaging resources.^32-34

Although deep learning–based SR methods have been broadly explored in medical imaging,^35-38 their application in HHUS requires particular attention to clinical plausibility. The approach evaluated in the present study prioritized preservation of diagnostically relevant plaque features, which might explain the observed improvements in diagnostic agreement without loss of specificity.

The improvement in identifying ultrasound-defined vulnerable plaque features should be interpreted with caution. Although SR-HHUS increased sensitivity compared with standard HHUS, sensitivity remained modest, at 63.2%, and this level of performance is not sufficient for SR-HHUS to serve as a stand-alone rule-out test for plaque vulnerability. In community practice, the more appropriate role of SR-HHUS might be as a triage support tool; clearer visualization of suspicious hypoechoic, irregular, or ulcerated plaques might prompt confirmatory vascular ultrasound, clinical risk-factor review, or referral when clinically indicated. Future models might improve vulnerable plaque assessment by integrating Doppler features, plaque texture quantification, clinical risk factors, and where available, multimodal imaging biomarkers.

Several limitations merit consideration. This study was conducted in a single region with a modest sample size, which might limit generalizability. The Carotid Plaque-Reporting and Data System was not applied, and plaque vulnerability was not defined by histology, magnetic resonance imaging, computed tomography angiography, or direct assessment of intraplaque hemorrhage, lipid-rich necrotic core, or fibrous cap thickness. Instead, we used B-mode ultrasound features associated with vulnerability, including hypoechogenicity, surface irregularity or ulceration, and plaque–lumen interface discontinuity. Therefore, the term “vulnerable plaque” in this study should be interpreted as ultrasound-defined vulnerable plaque features rather than comprehensive biological characterization of plaque instability. In addition, SR reconstruction was applied offline on a graphics processing unit–equipped workstation, and real-time inference on handheld ultrasound devices was not evaluated. Therefore, the present findings showed diagnostic feasibility but not yet real-time point-of-care implementation. Further work is needed to optimize model size, measure inference speed, and evaluate embedded or cloud-based deployment in community clinic workflows.

CONCLUSION

We developed an SR reconstruction algorithm based on Hyper-CycleGAN and applied it to HHUS imaging of carotid plaque in a community screening setting. Used as a postacquisition step on existing handheld devices, the model improved image sharpness, increased overall plaque detection, improved sensitivity for ultrasound-defined vulnerable plaque features, and strengthened agreement with reference ultrasound. Because sensitivity for ultrasound-defined vulnerable plaque features remained modest, SR-HHUS should be viewed as a triage support tool rather than a stand-alone diagnostic test for biological plaque vulnerability. Further validation in larger and more diverse settings is needed to assess real-time implementation, clinical workflow integration, and effect on referral decisions and preventive care.