Radiomics feature extraction from ultrasound B-mode images and radiofrequency signals of the carotid arterial wall: a feasibility study

doi:10.21203/rs.3.rs-5164876/v1

Background: Carotid ultrasound is largely used to assess arterial aging. Radiomics applied on ultrasound may allow characterizing wall ultrastructure and arterial aging, however to date its application to the intima-media complex is unexplored. The aim of this study is to investigate the feasibility and reproducibility of radiomic features extraction and selection in describing the common carotid wall ultrastructure.

Methods: Radiofrequency signals from 200 individuals in the Paris Prospective Study III cohort were used to extract RF and B-mode features. Feature stability across 3 frames from the same clip and 5 ROI sizes for each frame was evaluated by Intraclass Correlation analysis (threshold >0.50). Lasso L1 regression for chronological age prediction on ICC >0.50 features were used to analyze reproducibility and variation across frames and ROI sizes; 80/20 train-test split was used, with performance evaluated by MSE and R².

Results: Radiomic feature extraction was feasible in 190 out of 200 individuals. 48 features showed an ICC > 0.50. Feature selection for chronological age prediction showed consistent R² (0.09-0.14 train, 0.05-0.13 test) and MSE (32.3-34.3 train, 30.7-42.3 test) across frames. Fixed ROI sizes of 1.0mm and 1.2mm had comparable performance to a ROI size manually tailored to wall thickness: R² (0.07-0.15 train, 0.07-0.17 test), MSE (33.15-36.22 train, 33.07-34.83 test), consistently selecting substantially similar 3-6 features.

Conclusion: Radiomic ultrasound features describe carotid wall ultrastructure in a reproducible manner. While feature extraction and selection process are largely reproducible across frames, ROI size proved critical and thus must be carefully chosen.

Vascular ageing

Carotid radiomics

Arterial stiffness

Radio-frequency ultrasound

Machine learning

Vascular ageing is marked by progressive structural and functional deterioration of blood vessels that occurs with ageing. These changes over time increase the risk of developing cardiovascular disease(1,2). Vascular ageing is an inevitable physiological process; however, the rate and extent of this vascular deterioration can vary considerably among certain individuals. The Early vascular ageing (EVA) exhibits structural and functional arterial changes typically associated with older age, despite being chronologically younger(3,4). Conversely, in supernormal vascular ageing (SUPERNOVA) individuals demonstrate remarkable preservation of vascular health and function, with their arteries appearing physiologically younger than their chronological age. In recent years, there has been an increased evidence on the role of vascular ageing biomarkers in predicting cardiovascular events(5,6).

The current assessment of vascular ageing may involve utilizing biomarkers for atherosclerosis, such as the Calcium Score (CAC) or plaque analysis , as well as biomarkers for arteriosclerosis, including Pulse Wave Velocity (PWV) and Carotid Distensibility(7). These image-based assessments are typically performed using various imaging techniques, including Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and ultrasound (US). Thorough evaluations of both structural and functional vascular alterations are essential for assessing the cardiovascular risk associated with vascular ageing(8). Advanced imaging modalities like MRI and CT while effective, are expensive, complex and have low availability, thus not perfectly suitable for routine preventive risk assessments. In contrast, carotid US assessment is relatively low-cost, safe, widely available and able to capture both atherosclerotic and arteriosclerotic biomarkers. This makes it an optimal choice for routine preventive risk assessment. Among carotid ultrasound biomarkers, carotid plaque (9,10) and carotid stiffness (11) demonstrated predictive value for cardiovascular events, mostly stroke, on top of cardiovascular risk factors(12). Conversely Intima-media thickness (IMT) additive predictive value is debated: While IMT is associated with cardiovascular risk factors and is useful for understanding response to treatment (13), its ability to enhance risk prediction beyond existing methods is limited (14). For this reason, there is a need for continuous improvement of methods and techniques aimed at developing new descriptors, in order to better capture the subtle structural variations in the carotid arteries(8,15). It is now well documented that radiomics, which involves extracting numerical data from images, can provide more detailed information than simple images and in some settings can be comparable to diagnostic assessment by radiologist (16,17). Multiple studies have validated the predictive power of radiomic features (quantitative information mostly from CT or MRI images) and correlating these with clinical outcomes(16). Especially, radiomics has proven impactful in oncology, in survival prediction for esophageal cancer patients(18), in predicting breast cancer heterogenicity(19) and in assessing both tumor recurrence (20), and disease free survival in advanced rectal cancer (21). Cardiovascular applications of radiomics have also shown promise, mostly applied on coronary CT scans (22,23), but also on carotid plaque US (24). For example, Huang et al. (2022) explored the relationship between radiomic plaque features and clinical symptoms(25), while Le et al. (2021) assessed CT angiography radiomics for stroke prediction(26). Liu et al. (2024) developed a radiomics nomogram for stroke risk in diabetes(27), and Van Engelen et al. (2014) demonstrated that plaque texture radiomics could predict vascular events(28). To date, the carotid IM complex has never been investigated by radiomics, because of small tissue region size, rending difficult feature extraction from standard B-mode images(29). However, a more comprehensive analysis of IMT ultrastructure may provide additional information on atherosclerosis development at an earlier stage of disease, before plaque development. Furthermore, it may convey information about smooth muscle cells and extracellular matrix organization within the media, which are key determinants of arterial stiffness (30). We hypothesized that combining B-mode with Radiofrequency (RF) signal radiomic analysis would provide complementary information for comprehensive characterization of the intima-media complex. B-mode imaging offers grayscale representation of tissue acoustic properties, morphological information, and interface detection, while RF signals provide raw acoustic data with preserved phase information, higher axial sensitivity and greater spatial resolution (31). In this study, we aimed to evaluate the feasibility and reproducibility of B-mode and RF radiomic features in characterizing the ultrastructure of the carotid vessel wall. A key focus was to assess the impact of the selected region-of-interest (ROI) size and the frame used on the extracted radiomic features.

Study Population

A feasibility study on a subset of 200 individuals selected from the Paris Prospective Study III (PPS3) cohort was conducted. A targeted sampling approach was implemented, oversampling individuals with Type 2 Diabetes (T2D) by up to 20% to capture more extreme phenotypes associated with the condition. Additionally, the presence of carotid plaques in approximately 20% of the sample was enforced, ensuring representation of vascular complications relevant to T2D. To ensure a comprehensive range of blood pressure values, the population was divided into thirds: one-third within the 25th to 75th percentile range (120-141 mmHg), one-third below the 25th percentile (<120 mmHg), and one-third above the 75th percentile (≥141 mmHg). Furthermore, individuals aged between 50 to 75 years were uniformly included, ensuring a consistent distribution across the age range.

PPS3 is an ongoing community based prospective observational study conducted in Paris, France(32). The study protocol was approved by the Ethics Committee of Cochin Hospital (Paris, France) and was registered on the World Health Organization International Clinical Trials Registry platform (NCT00741728) on 08/25/2008. A total of 10,157 men and women aged 50–75 years were enrolled, who underwent a comprehensive preventive medical check-up, after signing an informed consent form. The vascular US was performed using Esaote PICUS Machine, Genova, Italy (128 RF linear array transducer with 7.5MHz). The raw Radio-frequency data were preserved to facilitate in-depth analysis. The inclusion criteria required the visibility of the intima-blood interface in at least some part of the far wall of the right common carotid artery, in a clear reconstructed B-mode image. Further details are available in the publication by PPS3 study group et al(32).

Ultrasound Data Processing

First, we developed a graphical user interface (GUI) using MATLAB software (MathWorks, Inc., Massachusetts, USA, version 2022b) to process raw radiofrequency signals and to reconstruct and process B-mode images. Additionally, we identified 178 Radiomics features (see description below) to be calculated from the selected region of interest (ROI).

Building on previous work (33), RF signals were transformed into B-mode ultrasound images using standard techniques.

Radiomic Features

A total of 74 radiomic B-mode features and 104 radiomic radio-frequency (RF) features were evaluated with the GUI. The B-mode features encompassed 1) First-Order Statistics(34) 2) Higher Order Textural Features (35–37)3) transform-based wavelet Features(38) 4) Fractal analysis features(39,40). Similarly, the RF features comprised(41) 1) Time series Features computed individually for each RF time series within the Region of Interest (ROI), with the mean value computed on 30 frames to derive time domain characteristics(42). 2) Frequency Domain Features involved Fourier transform to acquire the frequency spectrum, followed by straight-line fitting on the normalized spectrum(43,44). Furthermore, 3) Nakagami Distribution was utilized to extract the M parameter from the Nakagami distribution mean diagram (NDM) parametric map(43,45). 4) Spectral features(44). 5) Feature maps such as Direct energy attenuation diagram (DEA) and RF signal skewness intensity diagram (RF-I) were calculated and the texture analysis was applied to extract First-Order Statistics and Higher Order Textural Features from each map (43,46).

Data extraction settings

180 B-mode images (frames) were obtained from every original 6-second acquisition (a 128 radiofrequency lines multiarray with a depth of 4 cm captured at 30 frames per second). The Region of Interest was manually selected from the B-mode image capturing Intima Media complex on the far wall of the right carotid artery using a rectangular bounding box. Three end-diastolic frames from each patient were selected. For each frame, four ROI sizes (1 mm, 1.2 mm, 1.4 mm, and 1.6 mm) were extracted from the same location, with the Bounding Box centered on the smoothest section of the far wall to ensure optimal visualization of the Intima-blood interface (as depicted in Fig. 1). The Bounding Box encompassed the blood intima interface with minimal blood lumen on one side and the adventitia on the other. Initially, the bounding box was set at 1mm, gradually expanding by 0.2 mm towards the adventitia side while maintaining its position, in order to obtain the four different ROI sizes. Additionally, a fifth ROI size, termed the Variable ROI was introduced, which is the most suitable size among the four, precisely covering the Intima-Media (IM) complex (visually selected). Once extracted with the GUI, the features were normalized before performing feature selection.

Statistical analysis and Feature engineering

Descriptive statistics for population variables are presented as mean ± standard deviation (SD) or as counts (n) and percentages (%). First, we evaluated the feature stability across the three frames of the same clip and 5 ROI sizes of each frame by applying Intraclass Correlation (ICC) analysis with threshold of ICC > 0.50. We applied a two-way mixed effects model to calculate absolute agreement, treating ROI sizes as fixed effects and individuals as random effects (44). Second, we investigated the reproducibility across the three frames of the subset features with ICC > 0.50, for the prediction of chronological age (a proxy of vascular ageing), by applying least absolute shrinkage and selection operator (Lasso - L1 regularization) regression for feature selection (47). The following metrics were compared: model performance for age prediction by mean square error (MSE) and R², and number and type of selected features. Those metrics were calculated from four datasets: the three containing the features extracted by three selected frames and one containing their median values, using the variable ROI size. Internal validation was tested by 80/20 split sample technique.

Thirdly, the impact of variation in ROI size on model performance and selected features for chronological age prediction was also investigated by Lasso L1 regression. The following metrics were compared: model performance for age prediction by mean square error (MSE) and R², and number and type of selected features. Those metrics were calculated from the 5 datasets containing the median value of each feature for the three frames for 5 ROI sizes (1.0 mm, 1.2 mm, 1.4 mm, 1.6 mm and Var ROI). The internal validation was tested by 80/20 split sample technique. Additionally, we performed sensitivity analyses by applying Minimum Redundancy Maximum Relevance (MRMR) and Stepwise feature selection methods instead of Lasso L1, to validate the stability and reproducibility of the results with other feature selection techniques (see workflow diagram figure 2). The analyses were carried out using RStudio version 2023.9.0.463 (Boston, MA), utilizing glmnet, mlr ,caret , dplyr, mRMRe, e1071 and tidyverse packages.

Characteristics of the study population

Out of the initial cohort of 200 individuals, 10 were excluded because the reconstructed images did not display clearly the IMT complex in any frame, thus the analysis was run in 190 individuals. There were 48.42% (92) women, 40% (77) hypertensives, 15.78% (30) with the presence of carotid plaques, and 20% (37) diabetic individuals, with a mean IMT of 631 μm ± 108 standard deviation (std) and mean age of 59.47 years (Max 74.07- Min 50.01). The baseline characteristics of study population are mentioned in Table 2 and the carotid geometric and mechanical properties are mentioned in Table 3.

Feature stability across the frames

Feature showing ICC > 0.50 were 73 when ROI size was 1 mm, 72 when ROI size was 1.2 mm, 65 when ROI size was 1.4 mm, 60 when ROI size was 1.6 mm, and the variable ROI (which was used as a reference) yielded 48 features. Notably, all 48 features with ICC > 0.50 when ROI size was variable were also part of the sets identified by the fixed ROI sizes (See Table S1 in the Supplementary file). These features included 40 B-mode features (5 first order, 21 higher order, 9 wavelet transform features) and 8 RF features (5 first order, and 3 time series features).

Reproducibility of features across Frames

L1 regularization demonstrated consistent R2 (0.09-0.14 in the train set, 0.05-0.13 in the test set) and Mean square error values (32.3-34.3 in the train set, 30.7-42.3 in the test set) across the 4 datasets. Selected features varied in number from 2 to 10, but feature type and class were similar across frames. Selected features included B-mode first-order and higher-order features, as well as RF first order features (Table 4). Similar and consistent model performance was observed with different feature selection models (MRMR and stepwise regression, in Table S2 of Supplementary file).

Variability between ROI sizes:

Model performance for chronological age prediction varied considerably across different ROI sizes. The best performance was obtained by the Var ROI size, as expected; the model failed to select any feature in the 1.6 ROI size dataset. Overall MSE values ranged from 33.15-36.22 in the train to 33.07-34.83 in the test set and R² values ranged from 0.07-0.15 in the train set to 0.07- 0.17 in the test dataset. Interestingly, model performance and feature selected in ROI size 1.0 and 1.2 datasets were similar to variable ROI size (table 5). Furthermore, features selected in in ROI size 1.0 and 1.2 datasets were more numerous than those in Var ROI size, but type and class were similar. Consistent and similar results were obtained with different feature selection models (MRMR and stepwise regression, in Table S2 of Supplementary file).

Radiomic analysis is increasingly utilized in carotid US, primarily for plaque analysis with findings typically derived from B-mode imaging(24). This study is likely the first to apply radiomics in two novel ways: first, by focusing on the intima-media complex to characterize wall ultrastructure rather than solely on plaque; and second, by utilizing both B-mode images and raw radiofrequency signals, which contain richer spatial information than B-mode alone (26). While combined use of RF and B-mode US radiomic features has been applied in oncology, its application in cardiovascular radiomics remains unexplored. Incorporating RF-based radiomic features has improved accuracy in cancer detection and characterization compared to B-mode alone in breast cancer (48,49). The main objective of our study was to assess the feasibility and reproducibility of extracting radiomic features from the IM complex using both RF and B-mode US data. In terms of feasibility, RF and B-mode feature extraction was successfully performed on 190 out of 200 clips. The only clips where feature extraction was not feasible were those in which the IMT was not visible at all.

Once feasibility was established, we investigated issues related to ROI delineation, which is crucial for feature extraction integrity. Indeed, inaccurate ROI delineation could lead to incomplete or inaccurate representation of the target anatomical structure (IM complex). In particular, we investigated the impact of the cardiac cycle frame (temporal aspect) and of the size of the region of interest (spatial aspect) on the radiomic features.

First, our analysis of frame variability sought to recognize stable and reproducible features across 3 end-diastolic frames. By intraclass correlation coefficient we identified a robust set of features consistently present irrespective of the frame and ROI size (from B-mode first order, higher order, wavelet, RF first order, and time series classes), indicating their reproducibility in characterizing IM complex. We further investigated frame variability by testing radiomic feature capability of describing the process of ageing at the level of the IM complex, by using chronological age as an outcome in different variable selection models. The consistent selection of the same feature types across all frames (type of features) indicated that 1) they are likely to describe actual IM complex properties 2) choice of frame does not significantly impact the analysis, allowing for flexibility in frame selection without compromising the integrity of the results. The selected features from B-mode first order features describe intensity and contrast, reflecting IM complex density. Higher order features reveal texture and heterogeneity, indicating tissue structure. Wavelet features capture multi-scale structural changes, detecting alterations in the IM complex(50). RF first order features provide insights into acoustic properties, revealing tissue mechanical characteristics.

Second, we evaluated whether a fixed, automatically chosen, ROI size could provide results similar compared to a tailored ROI size manually drawn by an operator. Our results show that fixed ROI sizes of 1.0 or 1.2 mm, but not 1.4 and 1.6 mm, provide similar results in terms of feature reproducibility and selection compared to the variable ROI. This suggests that the choice of ROI size is critical in optimizing the extraction of relevant features, highlighting the importance of selecting an appropriate size tailored to the anatomical characteristics of the IM complex. Inclusion of portions of the adventitia in the ROI significantly alter the radiomic feature profile.

We acknowledge some limitations in this study. First, since this is a pilot investigation, it is conducted in a small highly selected population sample. Second, though internal validation has been tested by split sample technique, external validation has not been demonstrated yet. We plan to validate our findings in an external dataset that includes a broader age range and diverse ethnicities. Third, feature extraction has been performed only in end-diastolic frames: exploring the potential impact of different cardiac cycle phases on feature extraction could provide additional insights. Fourth, clinical relevance of the extracted features as well as potential confounding factors such as comorbidities, lifestyle factors, and environmental exposures is beyond the scope of this article, but needs to be investigated in future studies.

Our study demonstrates the feasibility and reproducibility of radiomic US features in characterizing carotid wall ultrastructure, from IM complex using both Rf and B-mode US data, showing minimal sensitivity to variations in frame selection. However, ROI size significantly affects feature extraction, highlighting the importance of precise ROI delineation in radiomics research. Building on these methodological improvements, we aim to expand radiomics applications in vascular health as novel biomarkers of vascular ageing

CAC: Calcium Score

CT: Computed Tomography

EVA: Early vascular ageing

ICC: Intraclass Correlation

IM: Intima-Media

IMT: Intima-Media Thickness

LASSO: Least Absolute Shrinkage and Selection Operator

MRI: Magnetic Resonance Imaging

MRMR: Minimum Redundancy Maximum Relevance

MSE: Mean Square Error

PWV: Pulse Wave Velocity

PPS3: Paris Prospective Study III

RF: Radiofrequency

ROI: Region-of-interest

SD: Standard deviation

SUPERNOVA: Supernormal Vascular Ageing

T2D: Type 2 Diabetes

US: Ultrasound

VAR: Variable

Ethics Approval and Consent to Participate

The study protocol was approved by the Ethics Committee of Cochin Hospital (Paris, France) and was registered on the World Health Organization International Clinical Trials Registry platform (NCT00741728) on 08/25/2008. A total of 10,157 men and women underwent a comprehensive preventive medical check-up, after signing an informed consent form.

Consent for Publication

All authors read and approved the final manuscript for publication.

Availability of Data and Materials

The data used in the current study are not publicly available due privacy issues but it will be made available on reasonable request. All data analyzed in this study are presented within the paper and supplementary material.

Competing Interests

The authors declare no competing interests.

Funding:

This work was supported by research grant from the European Commission Marie Skłodowska-Curie Actions PhD program: MINDSHIFT (grant number 954798, website: http://www.eumindshift.eu). The PPS3 (Paris Prospective Study III) was supported by grants from The National Research Agency (ANR), the Research Foundation for Hypertension (FRHTA), the Research Institute in Public Health (IRESP) and the Region Ile de France (Domaine d’Intérêt Majeur), and the H2020 ESCAPENET research program.

Authors’ Contributions:

MJ design and development of GUI, methodology, data analysis, interpretation of data and drafting the manuscript. FP, EB, FF worked on the design and development of the GUI, interpretation of data and critical review of the manuscript. HK data collection physician for PPS III study. XJ, JP conception, design and development of the PPS III study. PB, JP, RM conception of the research, interpretation of results and critical review of the manuscript.

Boutouyrie P, Chowienczyk P, Humphrey JD, Mitchell GF. Arterial Stiffness and Cardiovascular Risk in Hypertension. Circ Res. 2021 Apr 2;128(7):864–86.
Laurent S. Defining vascular aging and cardiovascular risk. Journal of Hypertension. 2012 Jun;30:S3–8.
Nilsson PM, Boutouyrie P, Laurent S. Vascular Aging: A Tale of EVA and ADAM in Cardiovascular Risk Assessment and Prevention. Hypertension. 2009 Jul;54(1):3–10.
M Nilsson P. Early Vascular Ageing - A Concept in Development. Eur Endocrinol. 2015 Apr;11(1):26–31.
Bruno RM, Nilsson PM, Engström G, Wadström BN, Empana JP, Boutouyrie P, et al. Early and Supernormal Vascular Aging: Clinical Characteristics and Association With Incident Cardiovascular Events. Hypertension. 2020 Nov;76(5):1616–24.
Laurent S, Boutouyrie P, Cunha PG, Lacolley P, Nilsson PM. Concept of Extremes in Vascular Aging: From Early Vascular Aging to Supernormal Vascular Aging. Hypertension. 2019 Aug;74(2):218–28.
Li A, Yan J, Zhao Y, Yu Z, Tian S, Khan AH, et al. Vascular Aging: Assessment and Intervention. CIA. 2023 Aug;Volume 18:1373–95.
Jamthikar AD, Gupta D, Saba L, Khanna NN, Viskovic K, Mavrogeni S, et al. Artificial intelligence framework for predictive cardiovascular and stroke risk assessment models: A narrative review of integrated approaches using carotid ultrasound. Computers in Biology and Medicine. 2020 Nov;126:104043.
Mantella LE, Colledanchise KN, Hétu MF, Feinstein SB, Abunassar J, Johri AM. Carotid intraplaque neovascularization predicts coronary artery disease and cardiovascular events. European Heart Journal - Cardiovascular Imaging. 2019 Nov 1;20(11):1239–47.
Sillesen H, Sartori S, Sandholt B, Baber U, Mehran R, Fuster V. Carotid plaque thickness and carotid plaque burden predict future cardiovascular events in asymptomatic adult Americans. European Heart Journal - Cardiovascular Imaging. 2018 Sep 1;19(9):1042–50.
Vasan RS, Pan S, Xanthakis V, Beiser A, Larson MG, Seshadri S, et al. Arterial Stiffness and Long-Term Risk of Health Outcomes: The Framingham Heart Study. Hypertension. 2022 May;79(5):1045–56.
Van Sloten TT, Sedaghat S, Laurent S, London GM, Pannier B, Ikram MA, et al. Carotid Stiffness Is Associated With Incident Stroke. Journal of the American College of Cardiology. 2015 Nov;66(19):2116–25.
Willeit P, Tschiderer L, Allara E, Reuber K, Seekircher L, Gao L, et al. Carotid Intima-Media Thickness Progression as Surrogate Marker for Cardiovascular Risk: Meta-Analysis of 119 Clinical Trials Involving 100 667 Patients. Circulation. 2020 Aug 18;142(7):621–42.
Yeboah J, McClelland RL, Polonsky TS, Burke GL, Sibley CT, O’Leary D, et al. Comparison of Novel Risk Markers for Improvement in Cardiovascular Risk Assessment in Intermediate-Risk Individuals. JAMA. 2012 Aug 22;308(8):788.
Reesink KD, Spronck B. Constitutive interpretation of arterial stiffness in clinical studies: a methodological review. American Journal of Physiology-Heart and Circulatory Physiology. 2019 Mar 1;316(3):H693–709.
Rogers W, Thulasi Seetha S, Refaee TAG, Lieverse RIY, Granzier RWY, Ibrahim A, et al. Radiomics: from qualitative to quantitative imaging. The British Journal of Radiology. 2020 Apr 1;93(1108):20190948.
Van Griethuysen JJM, Lambregts DMJ, Trebeschi S, Lahaye MJ, Bakers FCH, Vliegen RFA, et al. Radiomics performs comparable to morphologic assessment by expert radiologists for prediction of response to neoadjuvant chemoradiotherapy on baseline staging MRI in rectal cancer. Abdom Radiol. 2020 Mar;45(3):632–43.
Wang J, Yu X, Zeng J, Li H, Qin P. Radiomics model for preoperative prediction of 3-year survival-based CT image biomarkers in esophageal cancer. Eur Arch Otorhinolaryngol. 2022 Nov;279(11):5433–43.
Tsarouchi MI, Vlachopoulos GF, Karahaliou AN, Vassiou KG, Costaridou LI. Multi-parametric MRI lesion heterogeneity biomarkers for breast cancer diagnosis. Physica Medica. 2020 Dec;80:101–10.
Bhardwaj D, Dasgupta A, DiCenzo D, Brade S, Fatima K, Quiaoit K, et al. Early Changes in Quantitative Ultrasound Imaging Parameters during Neoadjuvant Chemotherapy to Predict Recurrence in Patients with Locally Advanced Breast Cancer. Cancers. 2022 Feb 28;14(5):1247.
Cui Y, Wang G, Ren J, Hou L, Li D, Wen Q, et al. Radiomics Features at Multiparametric MRI Predict Disease-Free Survival in Patients With Locally Advanced Rectal Cancer. Academic Radiology. 2022 Aug;29(8):e128–38.
Cheng X, Dong Z, Liu J, Li H, Zhou C, Zhang F, et al. Prediction of Carotid In-Stent Restenosis by Computed Tomography Angiography Carotid Plaque-Based Radiomics. JCM. 2022 Jun 6;11(11):3234.
Dong Z, Zhou C, Li H, Shi J, Liu J, Liu Q, et al. Radiomics versus Conventional Assessment to Identify Symptomatic Participants at Carotid Computed Tomography Angiography. Cerebrovasc Dis. 2022;51(5):647–54.
Hou C, Li S, Zheng S, Liu LP, Nie F, Zhang W, et al. Quality assessment of radiomics models in carotid plaque: a systematic review. Quant Imaging Med Surg. 2024 Jan;14(1):1141–54.
Huang Z, Cheng XQ, Liu HY, Bi XJ, Liu YN, Lv WZ, et al. Relation of Carotid Plaque Features Detected with Ultrasonography-Based Radiomics to Clinical Symptoms. Transl Stroke Res. 2022 Dec;13(6):970–82.
Le EPV, Rundo L, Tarkin JM, Evans NR, Chowdhury MM, Coughlin PA, et al. Assessing robustness of carotid artery CT angiography radiomics in the identification of culprit lesions in cerebrovascular events. Sci Rep. 2021 Feb 10;11(1):3499.
Liu Y, Kong Y, Yan Y, Hui P. Explore the value of carotid ultrasound radiomics nomogram in predicting ischemic stroke risk in patients with type 2 diabetes mellitus. Front Endocrinol. 2024 Apr 19;15:1357580.
Van Engelen A, Wannarong T, Parraga G, Niessen WJ, Fenster A, Spence JD, et al. Three-Dimensional Carotid Ultrasound Plaque Texture Predicts Vascular Events. Stroke. 2014 Sep;45(9):2695–701.
Molinari F, Zeng G, Suri JS. A state of the art review on intima–media thickness (IMT) measurement and wall segmentation techniques for carotid ultrasound. Computer Methods and Programs in Biomedicine. 2010 Dec;100(3):201–21.
Lacolley P, Regnault V, Segers P, Laurent S. Vascular Smooth Muscle Cells and Arterial Stiffening: Relevance in Development, Aging, and Disease. Physiological Reviews. 2017 Oct 1;97(4):1555–617.
Hu R, Singla R, Deeba F, Rohling RN. Acoustic Shadow Detection: Study and Statistics of B-Mode and Radiofrequency Data. Ultrasound in Medicine & Biology. 2019 Aug;45(8):2248–57.
on behalf of the PPS3 Study Group, Empana JP, Bean K, Guibout C, Thomas F, Bingham A, et al. Paris Prospective Study III: a study of novel heart rate parameters, baroreflex sensitivity and risk of sudden death. Eur J Epidemiol. 2011 Nov;26(11):887–92.
Standard B-Mode Ultrasound Measures Local Carotid Artery Characteristics as Reliably as Radiofrequency Phase Tracking in Symptomatic Carotid Artery Patients - Ultrasound in Medicine and Biology [Internet]. [cited 2024 Aug 21]. Available from: https://www.umbjournal.org/article/S0301-5629(15)00476-7/abstract
Ariyoshi K, Okuya S, Kunitsugu I, Matsunaga K, Nagao Y, Nomiyama R, et al. Ultrasound analysis of gray‐scale median value of carotid plaques is a useful reference index for cerebro‐cardiovascular events in patients with type 2 diabetes. J of Diabetes Invest. 2015 Jan;6(1):91–7.
Sim Y, Lee SE, Kim EK, Kim S. A Radiomics Approach for the Classification of Fibroepithelial Lesions on Breast Ultrasonography. Ultrasound in Medicine & Biology. 2020 May;46(5):1133–41.
Allison JW, Barr LL, Massoth RJ, Berg GP, Krasner BH, Garra BS. Understanding the process of quantitative ultrasonic tissue characterization. RadioGraphics. 1994 Sep;14(5):1099–108.
Jong Kook Kim, Hyun Wook Park. Statistical textural features for detection of microcalcifications in digitized mammograms. IEEE Trans Med Imaging. 1999 Mar;18(3):231–8.
Arivazhagan S, Ganesan L. Texture classification using wavelet transform. Pattern Recognition Letters. 2003 Jun;24(9–10):1513–21.
Al-Kadi OS, Watson D. Texture Analysis of Aggressive and Nonaggressive Lung Tumor CE CT Images. IEEE Trans Biomed Eng. 2008 Jul;55(7):1822–30.
Alic L, Niessen WJ, Veenland JF. Quantification of Heterogeneity as a Biomarker in Tumor Imaging: A Systematic Review. Hatzis C, editor. PLoS ONE. 2014 Oct 20;9(10):e110300.
Van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research. 2017 Nov 1;77(21):e104–7.
Zheng Q, Lin C, Xu D, Zhao H, Song M, Ou D, et al. A Preliminary Study on Exploring a potential Ultrasound Method for Predicting Cervical Cancer. J Cancer. 2022;13(3):793–9.
Xiao T, Shen W, Wang Q, Wu G, Yu J, Cui L. The detection of prostate cancer based on ultrasound RF signal. Front Oncol. 2022 Dec 12;12:946965.
Shams E, Karimi D, Moussavi Z. Bispectral analysis of tracheal breath sounds for Obstructive Sleep Apnea. In: 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society [Internet]. San Diego, CA: IEEE; 2012 [cited 2024 Feb 12]. p. 37–40. Available from: http://ieeexplore.ieee.org/document/6345865/
Tsui PH, Zhou Z, Lin YH, Hung CM, Chung SJ, Wan YL. Effect of ultrasound frequency on the Nakagami statistics of human liver tissues. Lee WN, editor. PLoS ONE. 2017 Aug 1;12(8):e0181789.
Wang Q, Dong Y, Xiao T, Zhang S, Yu J, Li L, et al. Prediction of programmed cell death protein 1 in hepatocellular carcinoma patients using radiomics analysis with radiofrequency-based ultrasound multifeature maps. BioMed Eng OnLine. 2022 Dec;21(1):24.
Tibshirani R. Regression Shrinkage and Selection Via the Lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology. 1996 Jan 1;58(1):267–88.
Klimonda Z, Karwat P, Dobruch‐Sobczak K, Piotrzkowska‐Wróblewska H, Litniewski J. Assessment of breast cancer response to neoadjuvant chemotherapy based on ultrasound backscattering envelope statistics. Medical Physics. 2022 Feb;49(2):1047–54.
Tadayyon H, Sannachi L, Gangeh MJ, Kim C, Ghandi S, Trudeau M, et al. A priori Prediction of Neoadjuvant Chemotherapy Response and Survival in Breast Cancer Patients using Quantitative Ultrasound. Sci Rep. 2017 Apr 12;7(1):45733.
Golemati S, Lehareas S, Tsiaparas NN, Chatziioannou A, Nikita KS, Perrea DN. Multiresolution features of carotid artery wall and plaque toward identifying vulnerable asymptomatic cases from B-mode ultrasound. In: 2013 IEEE International Ultrasonics Symposium (IUS) [Internet]. Prague, Czech Republic: IEEE; 2013 [cited 2024 Jul 17]. p. 872–5. Available from: http://ieeexplore.ieee.org/document/6725296/

Table 1 Summary of Extracted Radiomic Features and their Classes.

Feature Extraction Method	Feature Type	Feature Class	Total Number (n=count)	Features
B-mode	First Order Features		7	Mean value Skewness Kurtosis Entropy Energy Median value Stand deviation
	Higher Order Features (texture features)	Grey Level Distance Matrix (GLDM)	4	Mean entropy Entropy Contrast Angular second moment
		Neighboring Grey Level Dependence Matrix (NGLDM)	5	Coarseness Contrast Busyness Complexity Strength
		Grey Level Size Zone Matrix (GLSZM)	8	Small Zone Emphasis Large Zone Emphasis Gray-Level Non-uniformity Zone-Size Non-uniformity Zone Percentage Gray-Level Variance Zone-Size Variance Zone entropy
		Grey Level Run Length Matrix (GLRLM)	8	Short Run Emphasis Long Run Emphasis Gray-Level Nonuniformity Run-Length Nonuniformity Run Percentage Gray-Level Variance Run-Length Variance Run entropy
		Gray Level Dependence Matrix *	5	Small number emphasis (finess index) large number emphasis (coarsness index) Number non-uniformity Entropy DCENT entropy
		Averaged features across 4 directions	10	Mean homogeneity Mean energy Mean local homogeneity Mean Autocorrelation Mean Correlation Mean Dissimilarity Mean sum average Mean entropy Mean Contrast Mean variance
	Wavelet Transform	Wavelet features	24	MeanVett1- 12 (Mean) SdVett 1-12 (standard deviation)
	Fractal Analysis	Fractal Analysis features	3	FD average FD standard deviation FD lacunarity
Radio-frequency	Frequency domain	Spectral Features	12	spectral slope spectral Intercept Mid Band Fit S1 parameter S2 Parameter S3 Parameter S4 parameter signal power Spectral centroid Spectral Bandwidth spectral flatness Cress factor
	Time Domain Features	Time Series	5	Kurtosis cross zero count cross zero SD Peaks Fussy Entropy
	Nagakami distribution	Nakagami distribution mean diagram	1	M parameter
	Spatial Features	Skewness of spectrum difference Map	43	First Order Features* Grey Level Distance Matrix* Grey Level Size Zone Matrix* Grey Level Run Length Matrix * Gray Level Dependence Matrix * Averaged features across 4 directions*
		Direct Energy attenuation map	43	First Order Features* Grey Level Distance Matrix* Grey Level Size Zone Matrix* Grey Level Run Length Matrix* Gray Level Dependence Matrix * Averaged features across 4 directions*

Data are expressed as total number counts including both RF and B-mode based features (n=counts). * Indicates the same set of features described in the B-mode feature family.

Table 2 Summary of baseline characteristics of the population

Variables	Missing data	Overall Population(n=190)
Sex (Female) (n, %)	0	92 (48.4)
Smokers (n, %)	0	34 (17.8)
Diabetes (n, %)	0	37(19.4)
Hypertensive (n, %)	0	77(40.5)
Antidiabetic Drug Users (n, %)	1	21(11.05)
Antihypertensive Drug Users (n, %)	0	40(21.05)
Lipid Lowering Drug Users (n, %)	0	40(21.05)
Age (years)	0	59.5 ± 6.2
BMI (kg/m²)	0	25.2 ± 3.3
Mean Blood Pressure (mmHg)	0	94 ±11
SBP (mmHg)	0	132 ± 18
DBP (mmHg)	0	76 ± 9
Heart Rate (bat/min)	0	64 ± 10
HDL (mg/dL)	0	59.1 ±1 5.6
LDL (mg/dL)	0	142.9 ± 34.3
Cholesterol (mg/dL)	0	222.6 ± 38.4

Data for categorical variables are expressed as total number counts (n=counts) and percentage of total and for continuous variables as Mean ± SD (standard deviation).

Table 3 Summary of carotid geometric and mechanical characteristics

Carotid Variables	Missing data	Overall Population(n=190)
Presence of carotid Plaque (n, %)	0	30 (15.7)
Distension (μm)	3	359.88 ± 121
External diastolic Diameter (mm)	0	7.17 ± 0.73
Compliance (m²/kPa) Distensibility coefficient (kPa-1*10-3)	3 3	0.58 ± 0.23 21.43 ± 8.63
Carotid Pulse Wave velocity (m/s)	3	7.47 ± 1.57
IMT (um)	0	631 ± 108.6

Data for categorical variables are expressed as total number counts (n=counts) and percentage of total and for continuous variables as Mean ± SD (standard deviation).

Table 4 Results of reproducibility of features across frames.

Frame No.	Roi size (mm)	Model	Train MSE	Train R²	Test MSE	Test R²	Feature selected	Feature Name (Feature class)	Feature Type
1	Var	Lasso L1	34.36	0.09	40.28	0.05	4	Busyness and complexity (NGLDM), Gray-Level Nonuniformity (GLRLM), SdVett_5 (Wavelet transform)	B-mode higher order and Wavelet transform
2	Var	Lasso L1	32.31	0.14	38.31	0.13	10	Standard Deviation (First order), Entropy (GLDM), Contrast (NGTDM), Mean entropy (Avg features), Gray-Level Variance (GLRLM), meanVett_3, SdVett_5, SdVett_9 (Wavelet transform), and RF First order: Median (DEA) and Median (SSD)	B-mode first order, higher order, wavelet and RF first order
3	Var	Lasso L1	33.73	0.112	42.32	0.11	2	Gray-Level Nonuniformity (GLRLM), SdVett_9(Wavelet transform)	B- mode Higher order and Wavelet transform
Med	Var	Lasso L1	33.86	0.13	30.79	0.11	3	Complexity (NGTDM), Gray-Level Nonuniformity (GLRLM), SdVett_9 (Wavelet transform)	B- mode Higher order and Wavelet transform

Dependent Variable: Chronological age. The Mean square Error (MSE) is represented in years. *Feature selected are represented in count (number). Var is the optimal variable ROI used as reference. Where: DEA - Direct Energy attenuation map, GLDM- Grey Level Distance Matrix, GLRLM- Grey Level Run Length Matrix, NGLDM - Neighboring Grey Level Dependence Matrix and SSD - Skewness of spectrum difference Map.

Table 5 Results of variability between ROI sizes.

Median	ROI size	Model	Train MSE	Train R2	Test MSE	Test R²	Feature Selected	Feature Name (Feature class)	Feature Type
Med	Var	Lasso	33.86	0.13	30.79	0.17	3	Complexity (NGLDM), Gray-Level Nonuniformity (GLRLM), SdVett_9 (Wavelet transform)	B-mode higher order and Wavelet transform
Med	1	Lasso	33.15	0.15	33.07	0.11	5	Coarseness (NGTDM), Mean Correlation (Avg features), Run length Non-uniformity (GLRLM), meanVett_3 (Wavelet transform), and RF First order: Standard deviation (DEA)	B-mode higher order, Wavelet transform and Rf first order
Med	1.2	Lasso	33.24	0.14	31.87	0.15	6	Entropy (First order), Coarseness (NGTDM), Mean variance (Avg features), SdVett_1 and SdVett_9 (Wavelet transform) and RF first order: Standard deviation (DEA)	B-mode first, higher order, Wavelet transform and Rf first order
Med	1.4	Lasso	36.22	0.07	34.83	0.07	3	Skewness (first order), Coarseness (NGTDM), Gray level non-uniformity (GLRLM)	B-mode first and higher order
Med	1.6	Lasso	-	-	-	-	0	-	-

Dependent Variable: Chronological age. The Mean square Error (MSE) is represented in years. *Feature selected are represented in count (number). Var is the optimal variable ROI used as reference. Where: DEA - Direct Energy attenuation map, GLDM- Grey Level Distance Matrix, GLRLM- Grey Level Run Length Matrix, NGLDM - Neighboring Grey Level Dependence Matrix and SSD - Skewness of spectrum difference Map.

No competing interests reported.

SupplementaryMaterialFS.docx

Radiomics feature extraction from ultrasound B-mode images and radiofrequency signals of the carotid arterial wall: a feasibility study

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Results

Discussion

Conclusions

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1