Automated classification of atherosclerotic plaque from magnetic resonance images using predictive models☆
Introduction
Atherosclerosis is the most common cause of ischemic heart disease, which can lead to unstable angina, myocardial infarction (‘heart attack’), stroke, and sudden death. Studies indicate that plaque composition rather than the degree of stenosis or blood vessel narrowing is the key factor for predicting vulnerability to rupture or thrombosis (Fayad and Fuster, 2001, Fuster et al., 1999).
Several invasive methods have been used to identify plaque, including intravenous ultrasound (IVUS), angioscopy, intravascular MR, and thermography (Fayad and Fuster, 2001, MacNeill et al., 2003). These invasive tests are not appropriate for screening or serial examinations. Noninvasive imaging methods such as magnetic resonance imaging (MRI), ultrasound, or computed tomography (CT) may be used to obtain a much more accurate assessment of plaque burden, which facilitates the characterization of disease progression (Fayad et al., 2002, Fayad et al., 2004).
MRI is emerging as one of the dominant non-invasive methods for identifying and studying atherosclerotic plaque (Fayad and Fuster, 2001) (Corti et al., 2001, Fayad, 2003, Fayad et al., 2002, Fayad et al., 2004, Yuan et al., 2001a). A key advantage of MRI is that structures can be imaged using several contrast weightings. For example, T1-, T2-, and proton density-weighted images of the same anatomical tissue can be quite different, depending on the chemical components and structure of the tissue; hence, features not easily distinguished in one contrast weighting can show strong contrast in others (Cai et al., 2002, Shinnar et al., 1999, Yuan et al., 2001b).
Currently, atherosclerotic plaque is characterized by visual inspection by expert readers (Fayad and Fuster, 2001, Yuan et al., 2001b). This is a time-consuming and error-prone process, subject to several subjective biases. Classification accuracy is subject to inter- and intra-observer variability. Some automated tools have been developed to assist in the labeling process, including algorithms for aligning images (Kerwin et al., 2001, Suri and Laxminarayan, 2003), or outlining gross structure boundaries, such as the lumen or vessel wall boundaries (Chao et al., 2003, Mansard et al., 2004, Thomas et al., 2001).
Several groups have developed local texture segmentation algorithms for clustering sub-tissues, the most common method being a variant of K-means clustering (Xu et al., 2002, Yang et al., 2003, Yuan et al., 1999, Itskovich et al., 2004, Wolf et al., 2005). However, such unsupervised methods have high false-positives when used as classifiers and require a human operator to validate, adjust, and override the results. Clarke et al. (2003) extended this approach by developing a tissue-specific map, incorporating information from eight MR contrast weightings, and using a template matching approach to identify plaque components in endarterectomy specimens. This approach is similar to cluster analysis, with the important distinction that clusters are matched to idealized exemplars. In most cases, regression or predictive models have higher discrimination performance and are more robust (generalize to novel conditions) than template matching, as they operate on a less restricted solution space (Anderson et al., 2003).
Multiple contrast MR images offer a unique opportunity to improve feature classification accuracy as the number of image attributes is multiplied by the number of modalities. (The number of predictive variables that can be derived from combinations of these attributes is literally infinite.) However, human expert examination cannot take full advantage of this potential because of the complexity of the information that must be integrated. Unsupervised learning techniques – such as segmentation algorithms – can incorporate and summarize highly complex information, but do not take full advantage of what the expert knows from other sources of information such as anatomical cues or other diagnostic measurements. Supervised learning techniques, such as predictive models, capture relationships empirically, from known outcomes (in this case, post-operative histology) and thus combine statistical pattern classification with experience. Once a sample set of images has been labeled by an expert, a statistical model can be fit (using regression or optimization techniques) to exploit the collected evidence (Bishop, 1995).
In this study, we report the use of predictive modeling for plaque component classification from ex vivo MR images.
Section snippets
Methods
This section describes the methods used to prepare tissue samples and acquire MR images. Imaging was performed on a Bruker 9.4T MR system (Bruker Instruments, Billerica, MA). Seven human coronary artery specimens 1.28 cm long (1 from Pathobiological Determinants of Atherosclerosis in Young (PDAY) data library (Zieske et al., 2000) and 6 from the Institution's Autopsy Service) were imaged using a 10 mm birdcage coil. The tissue was then processed using customary histological procedures designed to
Results
Table 1 summarizes the performance of the RIPNet models based on two different statistical measures: the maximum Kolmogorov-Smirnov statistic and the Gini coefficient measure aspects of the ROC curve computed on the testing sample. An ROC curve is a graphical representation of the trade off between the false negative and false positive classification rates for every possible classification threshold, where for a given threshold:
Discussion
Automated plaque characterization (validated by histology, a consensus of experts, or long-term clinical outcomes) would establish a more accurate, consistent, and objective measure for both longitudinal and cohort studies. Any misclassification biases introduced by automation would be systematic; hence, comparisons of plaque burden in serial investigations on the same patient will yield a reliable measure of relative changes.
Without histology, classification of plaque from MR images remains
Acknowledgement
The authors wish to thank Halbert White, Ph.D. for his assistance with model estimation and statistical testing.
References (33)
MR imaging for the noninvasive assessment of atherothrombotic plaques
Magn. Reson. Imag. Clin. N Am.
(2003)- et al.
Closed contour edge detection of blood vessel lumen and outer wall boundaries in black-blood MR images
Magn. Reson. Imag.
(1999) - Anderson, R.W., Brotherton, T.W., Huizenga, J., 2003. Automated Methods and Systems for Vascular Plaque Detection and...
Neural Networks for Pattern Recognition
(1995)- et al.
Classification of human carotid atherosclerotic lesions with in vivo multicontrast magnetic resonance imaging
Circulation
(2002) - et al.
Detecting objects in image sequences using rule-based control in an active contour model
IEEE Trans. Biomed. Eng.
(2003) - et al.
Quantitative assessment of carotid plaque composition using multicontrast MRI and registered histology
Magn Reson Med.
(2003) - et al.
Effects of lipid-lowering by simvastatin on human atherosclerotic lesions: a longitudinal study by high-resolution, noninvasive magnetic resonance imaging
Circulation
(2001) - et al.
Clinical imaging of the high-risk or vulnerable atherosclerotic plaque
Circ. Res.
(2001) - et al.
Computed tomography and magnetic resonance imaging for noninvasive coronary angiography and plaque imaging: current and potential future concepts
Circulation
(2002)
Magnetic resonance imaging and computed tomography in assessment of atherosclerotic plaque
Curr. Atheroscler. Rep.
Acute coronary syndromes: biology
Lancet
Modified elastic tissue-Masson trichrome stain
Stain Technol.
Quantification of human atherosclerotic plaques using spatially enhanced cluster analysis of multicontrast-weighted magnetic resonance images
Magn. Reson. Med.
A quantitative vascular analysis system for evaluation of atheroscleriotic lesion by MRI
Estimation of error rates in discriminant analysis
Technometrics
Cited by (11)
A review of computational methods applied for identification and quantification of atherosclerotic plaques in images
2016, Expert Systems with ApplicationsCitation Excerpt :After training, the built model can be used to classify new patterns not belong to the training set. With respect to identification of atherosclerotic plaque components, supervised classification has been proposed in various studies (Anderson et al., 2006; Clarke et al., 2006; Clarke et al., 2003; van Engelen et al., 2012; Hofman et al., 2006; van Engelen et al., 2012) to classify each pixel into a tissue type. Although the aforementioned studies have proved to be effective on identifying atherosclerotic plaque components, some drawbacks must be highlighted.
Knowledge Discovery in Biomedical Data: Theory and Methods
2013, Methods in Biomedical Informatics: A Pragmatic ApproachNoninvasive Cardiovascular Imaging in Rheumatoid Arthritis: Current Modalities and the Emerging Role of Magnetic Resonance and Positron Emission Tomography Imaging
2012, Seminars in Arthritis and RheumatismCitation Excerpt :In large arteries such as the carotids, HR-MRI can provide comprehensive information about plaque size, composition, morphology, and biology (88). Most importantly from the clinical point of view, HR-MRI can reliably detect and quantify the key histopathologic correlates of plaque vulnerability: the presence of a lipid-rich necrotic core, fibrous cap thickness, intraplaque hemorrhage, calcification, and thrombus, with high resolution and reproducibility (85,89–94). A close correlation between HR-MRI classification of the carotid atherosclerotic lesions and histologic findings, based on the American Heart Association classification, was confirmed in a study of 60 patients imaged before endarterectomy (95).
Rationale and design of dal-PLAQUE: A study assessing efficacy and safety of dalcetrapib on progression or regression of atherosclerosis using magnetic resonance imaging and 18F-fluorodeoxyglucose positron emission tomography/computed tomography
2011, American Heart JournalCitation Excerpt :Finally, DCE images are obtained on 1 slice after administration of 0.2 mmol/kg of gadolinium-based contrast agent.20 At the core laboratory, magnetic resonance images are analyzed at a dedicated workstation by an experienced observer for the presence or absence of plaque components.31,32 The vessel wall boundaries and outer and inner walls of the common carotid and descending abdominal aorta are manually traced to derive vessel dimensions and morphologies.
Inflammatory arthritis and heart disease
2018, Current Pharmaceutical Design
- ☆
This research was funded, in part, by an NIH Phase I SBIR Grant awarded to Dr. Anderson and the ISCHEM Corporation. Partial support was also provided by: NIH/NHLBI R01 HL71021, NIH/NHLBI R01 HL78667 (ZAF).
- ✠
In memory of our friend and colleague, Dr. Vitalii Itskovich. Dr. Vitalii Itskovich worked with Drs. Zahi A. Fayad and Valentin Fuster for many years in the area of atherosclerosis research conducted at the Mount Sinai School of Medicine. Vitalii's graduate work in the area of aortic wave velocity measurements provided groundwork for his noted accomplishments in the field of MR pulse sequencing. He was instrumental in the development of sequences useful for atherosclerotic plaque imaging including REX, IR-TFL-DIFF, and GRASP. His expertise in the field of MR image analysis led to the development of algorithms based on cluster analysis for of automating plaque segmentation. His contributions extended to diverse areas of research such as the characterization of aortic root atherosclerosis, detection of lesion regression and in disease progression in Marfan's syndrome. In September of 2003, Vitalii was diagnosed with a rare and particularly virulent form of gastric cancer. Vitalii vigorously researched his cancer and became very involved with his treatment plans outliving all expectations before succumbing at age 34 on a cold yet sunny New York City January day.