| 1 |
Saba et al. (94) (2015) |
SDL |
The developed an accurate system used CT modality for segmenting and quantifying the calcium regions. The cohort consisted of 75 subjects (one slice per subject). The algorithm used mean shift algorithm along with adaptive threshold mechanism. The performance used four metrics namely correlation coefficient (CC) between manual and automated method (CC=0.978 and PoM of 0.915), dice similarity (mean of 0.85 with SD of 0.085), Jaccard Index (mean of 0.747 with SD of 0.12), and polyline distance metric (mean of 0.195 with S of 0.177) |
| 2 |
Lakadir et al. (82) (2017) |
SDL |
CNN algorithms were used for the segmentation of the plaque area. The unique capabilities of DL for segmentation were discussed such as automatic feature extraction. The plaque characterization was implemented using 90,000 patches which were extracted from an image database. The results showed a correlation of 0.90 with clinical assessment suggesting DL approaches are superior |
| 3 |
Biswas et al. (79) (2018) |
SDL |
The review concludes that for the CVD/stroke risk assessment, the DL paradigms are superior to non-AI (conventional) paradigms. cIMT measurement was carried by using a two-stage DL system. Stage-I was for feature extraction by a convolution layer-based encoder. Stage-II consisted of ML-based regression that smoothed the LI and MA borders as final output. The DB consisted of 396 B-mode US images. The system demonstrated a 20% improvement in the cIMT readings when compared to sonographer’s readings |
| 4 |
Sudha et al. (93) (2018) |
SDL |
The study adopted a two stage process where the stage-I consisted of a ROI extraction using deep learning CNN model, while stage-II consisted of border estimation of LI and MA using a deformable model (so-called snakes). The CNN model used a patch-based model as input for ROI extraction. The low data set consisted of 20 subjects giving a mean error difference of 0.08 mm between manual and automated process |
| 5 |
Azzorpardi et al. (78) (2019) |
SDL |
A novel Deep Neural Network-a fully automated tool for segmentation was developed, that deals with both LI and MA boundaries. A geometrically constrained function was proposed. The DICE measured for MA and LI regions was 0.962 and 0.925, respectively |
| 6 |
Wu et al. (84) (2019) |
SDL |
For diagnosing the CVD risk, the authors segmented the inner and outer walls of the carotid artery using DeepMed paradigm. The system demonstrated an accuracy of 90% |
| 7 |
Savaş et al. (91) (2019) |
SDL |
The study presented a CNN-based model for carotid artery wall classification using a database of 501 US scans. The ground truth labels for classification were taken by two specialists. The classification accuracy, sensitivity, and specificity of the model were 89.1%, 89%, and 88%, respectively |
| 8 |
Biswas et al. (81) (2020) |
SDL |
The study compared the patch-based solution in AI framework against the traditional (non-patch-based) method for segmentation of the walls of the ultrasound-based carotid artery. Post segmentation, the walls of the carotid artery was quantified jointly using plaque area (PA) and carotid IMT (cIMT). The system consisted of 250 carotid scans that give the cIMT error of 0.0935 mm (lowest of all previous methods) and PA error of 2.7939 mm2. The correlation coefficient’s between AI and GT for PA and cIMT was 0.89 and 0.99, respectively |
| 9 |
Jamthikar et al. (13) (2020) |
SDL |
Following was the three main finding of the paper: (a) there were two pathways mainly affected the atherosclerotic process, resulting in heart injury, (b) calculators that uses carotid US images performed better than the conventional calculators, (c) for CVD risk assessment in routine practice for RA patients using SDL AI methods |
| 10 |
Meshram et al. (83) (2020) |
SDL |
The study showed superior performance using UNet-based paradigm. For automated and semi-automated methods for plaque segmentation, the Dice-based performance metric yielded 0.48 and 0.83, respectively. Using the bounding box scheme with 5% error margin, the Dice was 0.79 and 0.80, respectively for UNet and dilated UNet paradigms |
| 11 |
Jain et al. (90) (2020) |
SDL |
The study presented a technique for localization of carotid artery in transverse B-mode ultrasound scans. The authors implemented a fast region convolutional neural network (FRCNN) system that combines regional proposal network (RPN) and object class detection network (PCDN). Using the correct bounding box, the highest IOU score obtained was 99%. The experiments were done for different number of epochs such as 30, 200, and 2000. The accuracies were highest for 2000 epochs showing the values as 89.91%, 89.71%, and 89.36% for K= 2, 5, and 10 respectively |
| 12 |
Groves et al. (106) (2020) |
HDL |
A comparison of mask R-CNN and UNet algorithms was carried out for automatically segmenting carotid artery (CA) and internal jugular vein (IJV). The mask RCNN produced a more accurate vascular segmentation and 3-D reconstruction of the vasculature. This yielded similar accuracy as the manually segmented CT scan. It enabled automatic analysis of the neck vasculature. A dataset consisted of 2439 images. The DICE scores generated for the CA and IJV were 0.90 and 0.88, respectively for masked R-CNN respectively. The DICE scores for UNet were 0.81 and 0.71, for the CA and IJV, respectively |
| 13 |
Tsakanikas et al. (107) (2020) |
HDL |
The carotid vessel segmentation was implemented using UNet-based CNN algorithm. Carotid atrial tree provided better results while the plaque tissue helped, in early detection treatment and prevention of carotid diseases. The UNet and morphological active contour were combined in a repetitive manner for segmenting the outer wall and carotid lumen. Using the MR image series obtained from TAXINOMSIS study, the system created a 3D meshed model by carotid bifurcation and smaller branches. The system showed an accuracy of 99.1% for lumen area and 92.6% for perimeter. Such models were applicable for computational fluid dynamics simulations |
| 14 |
Koktzoglou et al. (95) (2020) |
SDL |
The study introduced a strategy for optimizing the image quality in ungated 3T MRA for non-contrast extracranial carotid artery using DNN-based solution. The strategy took three minutes using single-shot radial sampling method, and it was benchmarked against 3-D filtering-based denoiser. The results of DL-based solution outperformed compared to 3D filtering method. Radial k-space sampling provided an increased arterial temporal signal-to-noise ratio (tSNR) by 107% and further improved image quality during 1-shot imaging. DL-based image processing gave a 24% and 195% increase when compared to original QISS score in arterial-to-background contrast and apparent contrast-to-noise ratio (aCNR) |
| 15 |
Xiao et al. (97) (2020) |
SDL (RF Signal) |
The study presented a DL-based model for segmentation of carotid vessel wall using RF signals. This was shown to be useful for understanding the mechanical properties of the carotid arteries, which indirectly measures the CVD risk. The DL-based methods showed a higher accuracy in tracking the wall motion. The DL method was compared against the block matching strategy. The performance was evaluated by using the displacement estimation error in Z and X direction. For Z-direction, the distance estimation error was 94.8% (DL method), 93.2% (block matching method). For X-direction, the distance estimation error was 94.2% (DL method) and 92.9% (block matching method) |
| 16 |
Biswas et al. (80) (2021) |
SDL |
The authors presented the review study that summarized the impact and evolution of cIMT/PA assessment using AI techniques. The authors compared four different segmentation techniques namely, traditional, semi-automated, ML, and DL-based. Further, ML/DL techniques were expressed in a mathematical way. The DL-based mechanisms are better due to automated feature extraction in the DL system |
| 17 |
Jain et al. (98) (2021) |
HDL |
This study uses the “Unseen AI” technique i.e., the training and testing datasets are from different ethnic groups. A four-layered UNet architecture were used for segmentation. The wall plaque area was measured for evaluation. The PE results for “Unseen AI” pair one were mean accuracy, DICE similarity, and CC with values 98.55, 78.38, and 0.80, respectively. For “Unseen AI” pair two, the values were 98.67, 82.49, and 0.87, respectively. The “seen AI” gave 99.01, 86.89, and 0.92, respectively. The DL-based models were validated for low atherosclerotic wall plaque segmentation by hypothesizing that “Unseen AI” lies in a very close proximity to “Seen AI” having a difference was almost less than 10%. The running time for the online system was almost less than one second. These DL-based methods can be used for CVD risk stratification |
| 18 |
Jain et al. (99) (2021) |
HDL |
The authors introduced the SegNet-UNet HDL architecture by applying it to a dataset with 970 ICA US images. Where, the SegNet was placed above the UNet for the HDL design. The performance were compared between five SDL/HDL architecture, namely UNet, UNet+, SegNet, SegNet-UNet, and SegNet-UNet+. The K10 model was applied in the study. The AUC values were 0.91, 0.911, 0.908, 0.905, and 0.898 (CE-loss) corresponding to UNet, UNet+, SegNet, SegNet-UNet, and SegNet-UNet+, respectively. Similarly, for DSC-loss the values were 0.883, 0.889, 0.905, 0.889, and 0.907, for the same order of models. The correlation values were 0.98, 0.96, 0.97, 0.98, and 0.97 for CE-loss models between AI-based PA and ground truth PA (GTPA). For DSC-loss models the values were 0.98, 0.98, 0.97, 0.98, and 0.98. It was concluded that SegNet-UNet outperformed other systems |
| 19 |
Zhou et al. (85) (2021) |
SDL |
The study demonstrated the usage of DL paradigm to segment the plaque in carotid longitudinal ultrasound scans, which was later used for TPA measurement. The authors showed high accuracy, low variability and suitability for (a) monitoring the response to therapy and (b) investigating new therapeutic methods for clinical trials. The two UNet models (M1 and M2) were trained with two GTs corresponding to two different observers. The performance evaluation for M1 and M2 models using Pearson correlation coefficient showed the values as 0.989, 0.987. The mean TPA difference when compared between UNet and manual segmentation ranged from 4.7 mm2 to 312.8 mm2. The mean segmentation time was 8.3±3.1 ms |
| 20 |
Zhou et al. (86) (2021) |
SDL |
The study presented DL paradigm for carotid plaque segmentation and TPA measurement, however, the study used a very small subset of data sets. A UNet++ ensemble algorithm was suggested in this study that used the 2D carotid US images. The gold standard used was a set of manually segmented images. The PE used consisted of the following metrics, namely, (I) Dice, (II) difference of area measurement (ΔTPA), (III) Pearson correlation coefficient, (IV) Bland-Altman plots, (V) intra-class correlation coefficient (ICC), and (VI) coefficient-of-variation. For SPARC dataset, DSC, ICC, CoV gave the values 83.3-85.7%, 0.996, and 7.54%, respectively, while for Zhongnan dataset, DICE and ICC was 88.6% and 0.985, respectively |
| 21 |
Ganitidis et al. (87) (2021) |
SDL |
The authors have introduced an interpretable classification paradigm for the risk assessment and plaque stratification using carotid US scans. The sampling was implemented by applying an ensemble learning scheme to maintain the balance between the symptomatic and asymptomatic classes. The primary ensemble-based models were built by using CNNs. Further, a six-layer deep CNN was used for automatic feature extraction process, followed by two fully connected layers. The DL system showed the three performance metrics such as area-under-the-ROC curve (AUC), sensitivity, and specificity having the values of 73%, 75% and 70%, respectively. The authors concluded that the clinical relevance was that the combination of DL paradigm with the interpretability methods and this facilitated the risk stratification process |
| 22 |
Mohannadi et al. (100) (2021) |
HDL |
The study adopted a DL-based technique that can be applied to perform semantic segmentation of intima-media complex, and further calculating the cIMT measurement. A fully automated encoder-decoder based prototype was used to overcome the lack of large size datasets as it used multi-image inputs providing a good learning to the system with different features. The encoder and decoder had stages namely, convolution, batch normalization, and parametric ReLU (PReLU). The performance evaluation was done by measuring F1-measure, precision, recall, DICE coefficient, JI having values of 79.92%, 81.18%, 82.06%, 74.23%, and 60.24%, respectively. From the obtained results, the authors concluded that the proposed encoder-decoder architecture out performed all other state-of-the-art work |
| 23 |
Latha et al. (101) (2021) |
HDL |
The authors used US images of carotid artery for IMT and plaque diameter. ML and DL approaches were used for recognizing symptomatic or asymptomatic plaque in a total of 361 images (202 normal images and 159 images with carotid plaque). The ML-based algorithm used were CART DT, RF, and logistic regression (LR), while DL-based models used were CNN, MobileNet, and CapsuleNet. The classification accuracy of CapsuleNet TL was 96.7%, unlike RF gave and accuracy of 91.41% |
| 24 |
Otgonbaatar et al. (102) (2021) |
HDL |
The study was an application in CVD risk stratification using CT angiography. Compared to filtered-back projection and hybrid iterative reconstruction, the DL-based reconstruction was superior demonstrating small vessels. This helped in blooming artifact reduction thereby improving the image quality |
| 25 |
Jain et al. (103) (2021) |
HDL |
The study adopted a DL-based solution using UNet and SegNet-UNet, while keeping the objective of speed, accuracy, and reliability during early detection and quantification of plaque lesions in CCA ultrasound scans. The system was benchmarked against AtheroEdge™ 2.0, demonstrating the accuracies of 93%, 94%, and 95%, respectively, corresponding to UNet, SegNet-UNet, and AtheroEdge™ 2.0 systems |
| 26 |
Ziegler et al. (88) (2021) |
SDL |
The study showed the usage of SDL model based on UNet, keeping the spirit of branch-level segmentation of carotid arteries that benefitted large-cohort investigations. The performance evaluation yielded Dice, Mathew, and true positive ratio of 0.80, 0.80, and 0.84, respectively |
| 27 |
Bortsova et al. (89) (2021) |
SDL |
The study used automated UNet-based DL solution for segmentation and intracranial carotid artery calcification (ICAC) volume measurement. While comparing against the manual methods, the sensitivity and PPV were 83.3%, and 88%, respectively. The authors showed the correlation between ICAC and incident stroke |
| 28 |
Zhu et al. (104) (2021) |
HDL |
The study used 3D residual-UNet DL approach for segmentation of lumen and wall in a diseased carotid artery. The identity mapping was done by using an optimal channel fitting structure. The strategies used for training the MRimages are patch-level and global level. Optimization was done to the pre-segmentation results, later cascaded with the patch-croped MRvolume data and trained for segmenting the carotid lumen and wall. The segmentation was reproduceable and showed the Dice of 0.84 and 0.74 for lumen and wall, respectively |
| 29 |
Wasih et al. (92) (2021) |
SDL |
The study presented two sets of models, namely automated RCNN and UNet for segmentation carotid artery, internal jugular vein from the transverse US scans of neck. The RCNN model was used for mask generation while UNet model was used for selection of the largest connected region for each class. The US models were validated using CT-based imaging. The performance was evaluated using Dice score which came out to be within two mm between US and CT |
| 30 |
Flores et al. (108) (2021) |
HDL |
This study presented a review using MI and AI for peripheral artery disease (PAD). Finally, the study discusses the potential areas for the future of PAD care and advanced solution such as analytics. The DL reconstruction techniques improves the image quality of brain CT angiography. The objective measurement and the subjective grading got improved when compared with filtered-back-projection and hybrid iterative reconstructions |
| 31 |
Luo et al. (96) (2021) |
SDL (Doppler US) |
The study developed a DL-based model to classify aortoiliac, femoropopliteal, and trifurcation disease in the US studies. This was then benchmarked against RF-based ML algorithm for classification of carotid stenosis in duplex US. The experienced physician was used for gold standard readings. The NN model used waveforms, pressures values, flow velocities, and plaque presence. AI was developed to automate the interpretation of these LEAD and carotid duplex ultrasound studies. The DL model obtained the performances in the form of accuracy of prediction of normal, aortoiliac disease, femoropopliteal disease, and trifurcation disease as 97%, 88.2%, 90.1%, and 90.5%, respectively. For internal carotid artery stenosis, the accuracies were classified as per the stenosis range 0-49%, 50-69%, > 70%, and 100% occlusion, having the accuracies of 99.2%, 100%, 100%, and 100%, respectively |
| 32 |
Rim et al. (166) (2021) |
SDL (CAC Score) |
The study showed that DL-based retinal photograph-derived CAC score that can be used as an alternative to CT scan-measured CAC in evaluating the cardiovascular events. The system RetiCAC showed a superior performance resulting an AUC of 0.742, when compared to single parameter models (Age: 0,705, glucose: 0.637) |
| 33 |
Park et al. (105) (2022) |
HDL |
The study proposed DL-UFV neural network-based on segmentation. The system combined the speckle tracking velocimetry and the speckle image velocimetry for vessel wall segmentation. The parameters measured were vascular stiffness and velocity field information of blood flow. The system improved biases in measurements of velocity, wall shear stress (WSS), and strain by 4.6-fold, 115.1-fold, and 22.2-fold, respectively. |
| 34 |
Jain et al. (161) (2022) |
HDL |
The study designed HDL models which was then benchmarked against the conventional SDL models. The HDL designed were Inception-UNet, Squeeze-UNet, and Fractal-UNet. The benchmarked SDL models were UNet, UNet+, UNet++ and UNet+++. The HDL models showed low memory, faster operations, and small training time of parameters. The coefficient of correlation metric provided 0.96, 0.96, 0.98, 0.95, 0.96, and 0.96 for CCA for seven SDL and HDL models respectively, whereas ICA resulted 0.99, 0.99, 0.98, 0.99, 0.98, 0.98 and 0.98 respectively. AUC for CCA images were 0.97, 0.969. 0.974, 0.969, 0.962, and 0.960 respectively, while for ICA images were 0.99, 0.989, 0.988, 0.989, 0.986, 0.989, and 0.988, respectively (P<0.001) |
