Introduction
Each year CVD accounted for 17.9 million deaths, worldwide [4]. There was a surge of ~25% in CVD mortalities from 2000 to 2016 and the majority (85%) were either due to ischemic heart disease or stroke. Low-and-middle-income countries with sociodemographic index <0.75 [5] reported higher CVD death rate [6,7]. Further, years of life lost (YLL) increased in CVD while global trends in total YLL decreased in other diseases [8]. The all-age disability-adjusted life-years (DALY) in CVD soared 6.5% while age-standardized DALY declined 16.9% from 2005 to 2015 [9]. Furthermore, financial stress raised incident CVD risk and all-cause mortality in destitute women and single-men [10,11]. The risk is more pronounced in diabetes [12] and hypertension [13]. To summarize, CVDs deteriorates the quality of life both health-wise and financially; it should, therefore, be identified at an earlier stage and treated [5].
Atherosclerosis, the process of narrowing arterial walls, is the underlying reason for CVD [14,15]. Expressed as a thickening on arterial walls, atherosclerosis leads to macrophage accumulation, necrotic core formation, fibrous-cap, and plaque buildup [16]. Plaque accumulates, perhaps, as a result of an inflammatory response owing to the progress in atherosclerosis [16]. Several complications of atherosclerosis; for example, plaque rupture, stenoses, obstruction to blood flow and embolism, lead to severe conditions in CVD [17]. Atherosclerosis appearing in early childhood and young are linked to CVD and remain silent (asymptomatic) for several years [18–20]. However, early-stage detection can prevent disease progression by suitable lifestyle modification and medical intervention [17]. Several studies have observed in the literature that arterial stiffness (or stiffness) and carotid intima-media thickness (IMT) are indicators of atherosclerosis from inception [21–23].
Arterial stiffness describe the rigidity of arterial walls (or loss of elasticity); is age-dependent [21], and augment as atherosclerosis progress [24–27]. Several pathological changes involved in the development of atherosclerosis are instigated in stiffness and both often coexist. Note that mature adults with depression had augmented IMT and stiffness values [28]. Further, arterial stiffness leads to elevated blood pressure [29–32] and in hypertensives, augmented stiffness indicate atherosclerotic CVD [33]. Importantly, elevated stiffness is associated with atherosclerotic cardiovascular events (CVE) [34], stroke [23,35], diastolic heart failure [33], hypertension [29,30,36], diabetes [37,38], obesity [39–41] and renal disease [42,43].
Several indices, pulse pressure (PP), compliance (C), distensibility (D), pulse wave velocity (PWV), Young’s modulus, augmented index (AIx), β-stiffness and Cardio-ankle vascular index (CAVI) are applied for the quantification of arterial stiffness. PP is the difference between systolic and diastolic pressure. Compliance is often described as the ratio of change in volume to change in pressure (\(C=\frac{\text{ΔV}}{\text{ΔP}}\)), and distensibility is viewed as the compliance per original volume (\(D=\frac{C}{V}\)) [44]. A brief description of various stiffness indices used in clinical studies is provided in Appendix A .
PWV, the most significant measure owing to its association with CVE, is defined as the velocity of pressure, diameter or flow-velocity wave. Often significant variation in elasticity and impedance is observed in arteries from proximal to peripheral [45]. This variation leads to amplification, reflection and change in wave shape along the vasculature attributed to composition, distending pressure, and muscle-tone of the arterial wall [46] [45][47] [48]. Further, the wave shape is dependent on age group [44,49,50], ageing [51,52], physical fitness [53,54], insulin produced [55,56], heart rate, body demographics and gender [57–59]. In short, pulse wave is indicative of hypertension [60], diabetes [56,61,62], aortic disintegration and heart failure [57,63]. From this point of perspective, the assessment of PWV is a promising approach for predicting atherosclerotic CVD.
Assessment of PWV includes invasive and non-invasive techniques. However, non-invasive techniques are popular owing to its clinical application. Clinically, PWV is estimated as the ratio of the distance between two points in distal arteries and time taken by the wave to cover this distance [64]. Based on the methodology used, non-invasive assessment techniques are either “two-site” (regional) or “one-site” (local). The two-site method includes traditional “foot-by-foot” method [64] while the one-site method requires multiple variables for assessment. Studies observed that branchial-ankle PWV (ba-PWV) and carotid-femoral PWV (cfPWV) are highly correlated with significant methods used in the assessment of PWV [65]. Acceptance of any particular method for clinical purposes was debated earlier [46,65]. However, cfPWV is considered “gold standard” due to its strong association with CVD [37,43] and recommendation by various committees [66]. Several challenges owing to the unstandardised nature of assessment are discussed in a later section. Further, 80% of the measured value of cfPWV and 10 m/s cutoff value is accepted clinically [67]. The one-site method includes image-based techniques that involve localisation and segmentation of the lumen region. Performance of these techniques depends on the appearance of the artery in a frame, operator experience, and equipment variability. However, machine learning algorithms proved efficient in localisation and segmentation of lumen and intima-media complex (IMC).
Machine learning systems extract patterns from raw-data and perform specific tasks. These systems are supervised or unsupervised based on the availability of labelled data [68]. During training, data is presented to the system and it produces an output vector based primarily on input and weights. An objective function measures error between the output vector and the target vector (desired output) [69]. The system then updates the weights to minimize this error [69]. This technique is repeated for small sets of data until the mean of the objective function stops decreasing [69]. Further, test-data is used to assess performance on a generalized data and validation is done for optimization [69]. The inspiration behind the application of DL framework is to obtain human-like perfectness in tasks like segmentation, object identification and classification. The application of DL in various types of applications, such as lumen characterisation, tumour segmentation and cell detection, has been a motivation for its application on the carotid artery [70].
Major deep learning architectures include the recurrent neural networks (RNN) [69], convolutional neural network (CNN) [69], deep belief networks (DBN) [70] and autoencoders [71,72]. RNN is a network wherein, the output of one layer is connected to the next layer. Partial outputs are applied as feedback along with the next input [69]. RNN is applied in speech and handwriting recognition. The key disadvantage is the problem of vanishing gradients observed in the backpropagation algorithm. Likewise, CNN is applied in object identification and classification. The hidden layer in CNN includes stacks of convolution layers, non-linear activation layers and pooling layers [69]. Moving from the input layer, the first few layers extract low-level features (edges) and “deeper” layers extract high-level features specific to the object [69]. This network is often trained using backward propagation [69]. Further, DBNs are applied in image and video recognition tasks. DBN utilizes unsupervised learning, based on the restricted Boltzmann machine (RBM) [70]. RBM is a two-layered network wherein the first (input) is called the visible layer and the second is called the hidden layer. In DBN, several RBMs are stacked: hidden layer of one RBM is connected to the visible layer of second RBM and so on. Each RBM is independently trained using unsupervised learning and backpropagation network is applied at the end layers [70]. Likewise, autoencoders also learn features from raw data in an unsupervised manner. This method offers efficient dimensionality reduction and denoising capability. Herein, unlabeled inputs are represented (encoded) using the most important features and the original image (input image) is reconstructed from the encoded image. Different types of autoencoders in practice are convolutional autoencoder, deep autoencoder and contractive encoder. The most important challenge faced by DL-based systems is the limited dataset. However, success using transfer learning methods have motivated research in using other methods, such as reinforced learning and autoencoders. Even so, DL architectures suffer due to overfitting and underfitting. Overfitting occurs when DL architecture gets trained to features more than what is required and underfitting is when too little data is provided, and the system is not modelled accurately. There are several methods that are effectively applied to surmount underfitting and overfitting.
This review summarizes state-of-the-art techniques used for the assessment of arterial stiffness. The article is organized into five main sections: Section 2. will begin by describing conventional techniques of PWV assessment and Section 3. will provide an insight into image-based techniques. Besides, Section 4. will investigate the application of DL-based systems in clinical diagnosis and the last section will discuss stiffness computation using DL-framework.