Harman S. Suri

Accurate cloud-based smart IMT measurement, its validation and stroke risk stratification in carotid ultrasound

This study presents AtheroCloud™ - a novel cloud-based smart carotid intima-media thickness (cIMT) measurement tool using B-mode ultrasound for stroke/cardiovascular risk assessment and its stratification. This is an anytime-anywhere clinical tool for routine screening and multi-center clinical trials. In this pilot study, the physician can upload ultrasound scans in one of the following formats (DICOM, JPEG, BMP, PNG, GIF or TIFF) directly into the proprietary cloud of AtheroPoint from the local server of the physicians office. They can then run the intelligent and automated AtheroCloud™ cIMT measurements in point-of-care settings in less than five seconds per image, while saving the vascular reports in the cloud. We statistically benchmark AtheroCloud™ cIMT readings against sonographer (a registered vascular technologist) readings and manual measurements derived from the tracings of the radiologist. One hundred patients (75 M/25 F, mean age: 68±11 years), IRB approved, Toho University, Japan, consisted of Left/Right common carotid artery (CCA) artery (200 ultrasound scans), (Toshiba, Tokyo, Japan) were collected using a 7.5MHz transducer. The measured cIMTs for L/R carotid were as follows (in mm): (i) AtheroCloud™ (0.87±0.20, 0.77±0.20); (ii) sonographer (0.97±0.26, 0.89±0.29) and (iii) manual (0.90±0.20, 0.79±0.20), respectively. The coefficient of correlation (CC) between sonographer and manual for L/R cIMT was 0.74 (P<0.0001) and 0.65 (P<0.0001), while, between AtheroCloud™ and manual was 0.96 (P<0.0001) and 0.97 (P<0.0001), respectively. We observed that 91.15% of the population in AtheroCloud™ had a mean cIMT error less than 0.11mm compared to sonographers 68.31%. The area under curve for receiving operating characteristics was 0.99 for AtheroCloud™ against 0.81 for sonographer. Our Framingham Risk Score stratified the population into three bins as follows: 39% in low-risk, 70.66% in medium-risk and 10.66% in high-risk bins. Statistical tests were performed to demonstrate consistency, reliability and accuracy of the results. The proposed AtheroCloud™ system is completely reliable, automated, fast (3-5 seconds depending upon the image size having an internet speed of 180Mbps), accurate, and an intelligent, web-based clinical tool for multi-center clinical trials and routine telemedicine clinical care.

Stroke Risk Stratification and its Validation using Ultrasonic Echolucent Carotid Wall Plaque Morphology

Stroke risk stratification based on grayscale morphology of the ultrasound carotid wall has recently been shown to have a promise in classification of high risk versus low risk plaque or symptomatic versus asymptomatic plaques. In previous studies, this stratification has been mainly based on analysis of the far wall of the carotid artery. Due to the multifocal nature of atherosclerotic disease, the plaque growth is not restricted to the far wall alone. This paper presents a new approach for stroke risk assessment by integrating assessment of both the near and far walls of the carotid artery using grayscale morphology of the plaque. Further, this paper presents a scientific validation system for stroke risk assessment. Both these innovations have never been presented before. The methodology consists of an automated segmentation system of the near wall and far wall regions in grayscale carotid B-mode ultrasound scans. Sixteen grayscale texture features are computed, and fed into the machine learning system. The training system utilizes the lumen diameter to create ground truth labels for the stratification of stroke risk. The cross-validation procedure is adapted in order to obtain the machine learning testing classification accuracy through the use of three sets of partition protocols: (5, 10, and Jack Knife). The mean classification accuracy over all the sets of partition protocols for the automated system in the far and near walls is 95.08% and 93.47%, respectively. The corresponding accuracies for the manual system are 94.06% and 92.02%, respectively. The precision of merit of the automated machine learning system when compared against manual risk assessment system are 98.05% and 97.53% for the far and near walls, respectively. The ROC of the risk assessment system for the far and near walls is close to 1.0 demonstrating high accuracy.

A Review on Atherosclerotic Biology, Wall Stiffness, Physics of Elasticity, and Its Ultrasound-Based Measurement

Functional and structural changes in the common carotid artery are biomarkers for cardiovascular risk. Current methods for measuring functional changes include pulse wave velocity, compliance, distensibility, strain, stress, stiffness, and elasticity derived from arterial waveforms. The review is focused on the ultrasound-based carotid artery elasticity and stiffness measurements covering the physics of elasticity and linking it to biological evolution of arterial stiffness. The paper also presents evolution of plaque with a focus on the pathophysiologic cascade leading to arterial hardening. Using the concept of strain, and image-based elasticity, the paper then reviews the lumen diameter and carotid intima-media thickness measurements in combined temporal and spatial domains. Finally, the review presents the factors which influence the understanding of atherosclerotic disease formation and cardiovascular risk including arterial stiffness, tissue morphological characteristics, and image-based elasticity measurement.

Comparative approaches for classification of diabetes mellitus data: Machine learning paradigm

BACKGROUND AND OBJECTIVE Diabetes is a silent killer. The main cause of this disease is the presence of excessive amounts of metabolites such as glucose. There were about 387 million diabetic people all over the world in 2014. The financial burden of this disease has been calculated to be about

Intra- and inter-operator reproducibility of automated cloud-based carotid lumen diameter ultrasound measurement

13,700 per year. According to the World Health Organization (WHO), these figures will more than double by the year 2030. This cost will be reduced dramatically if someone can predict diabetes statistically on the basis of some covariates. Although several classification techniques are available, it is very difficult to classify diabetes. The main objectives of this paper are as follows: (i) Gaussian process classification (GPC), (ii) comparative classifier for diabetes data classification, (iii) data analysis using the cross-validation approach, (iv) interpretation of the data analysis and (v) benchmarking our method against others. METHODS To classify diabetes, several classification techniques are used such as linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and Naive Bayes (NB). However, most of the medical data show non-normality, non-linearity and inherent correlation structure. So in this paper we adapted Gaussian process (GP)-based classification technique using three kernels namely: linear, polynomial and radial basis kernel. We also investigate the performance of a GP-based classification technique in comparison to existing techniques such as LDA, QDA and NB. Performances are evaluated by using the accuracy (ACC), sensitivity (SE), specificity (SP), positive predictive value (PPV), negative predictive value (NPV) and receiver-operating characteristic (ROC) curves. RESULTS Pima Indian diabetes dataset is taken as part of the study. This consists of 768 patients, of which 268 patients are diabetic and 500 patients are controls. Our machine learning system shows the performance of GP-based model as: ACC 81.97%, SE 91.79%, SP 63.33%, PPV 84.91% and NPV 62.50% which are larger compared to other methods.

Deep learning strategy for accurate carotid intima-media thickness measurement: An ultrasound study on Japanese diabetic cohort

Background Common carotid artery lumen diameter (LD) ultrasound measurement systems are either manual or semi-automated and lack reproducibility and variability studies. This pilot study presents an automated and cloud-based LD measurements software system (AtheroCloud) and evaluates its: (i) intra/inter-operator reproducibility and (ii) intra/inter-observer variability. Methods 100 patients (83 M, mean age: 68 ± 11 years), IRB approved, consisted of L/R CCA artery (200 ultrasound images), acquired using a 7.5-MHz linear transducer. The intra/inter-operator reproducibility was verified using three operator’s readings. Near-wall and far carotid wall borders were manually traced by two observers for intra/inter-observer variability analysis. Results The mean coefficient of correlation (CC) for intra- and inter-operator reproducibility between all the three automated reading pairs were: 0.99 (P < 0.0001) and 0.97 (P < 0.0001), respectively. The mean CC for intra- and inter-observer variability between both the manual reading pairs were 0.98 (P < 0.0001) and 0.98 (P < 0.0001), respectively. The Figure-of-Merit between the mean of the three automated readings against the four manuals were 98.32%, 99.50%, 98.94% and 98.49%, respectively. Conclusions The AtheroCloud LD measurement system showed high intra/inter-operator reproducibility hence can be adapted for vascular screening mode or pharmaceutical clinical trial mode.

Accurate Diabetes Risk Stratification Using Machine Learning: Role of Missing Value and Outliers

MOTIVATION The carotid intima-media thickness (cIMT) is an important biomarker for cardiovascular diseases and stroke monitoring. This study presents an intelligence-based, novel, robust, and clinically-strong strategy that uses a combination of deep-learning (DL) and machine-learning (ML) paradigms. METHODOLOGY A two-stage DL-based system (a class of AtheroEdge™ systems) was proposed for cIMT measurements. Stage I consisted of a convolution layer-based encoder for feature extraction and a fully convolutional network-based decoder for image segmentation. This stage generated the raw inner lumen borders and raw outer interadventitial borders. To smooth these borders, the DL system used a cascaded stage II that consisted of ML-based regression. The final outputs were the far wall lumen-intima (LI) and media-adventitia (MA) borders which were used for cIMT measurements. There were two sets of gold standards during the DL design, therefore two sets of DL systems (DL1 and DL2) were derived. RESULTS A total of 396 B-mode ultrasound images of the right and left common carotid artery were used from 203 patients (Institutional Review Board approved, Toho University, Japan). For the test set, the cIMT error for the DL1 and DL2 systems with respect to the gold standard was 0.126 ± 0.134 and 0.124 ± 0.100 mm, respectively. The corresponding LI error for the DL1 and DL2 systems was 0.077 ± 0.057 and 0.077 ± 0.049 mm, respectively, while the corresponding MA error for DL1 and DL2 was 0.113 ± 0.105 and 0.109 ± 0.088 mm, respectively. The results showed up to 20% improvement in cIMT readings for the DL system compared to the sonographers readings. Four statistical tests were conducted to evaluate reliability, stability, and statistical significance. CONCLUSION The results showed that the performance of the DL-based approach was superior to the nonintelligence-based conventional methods that use spatial intensities alone. The DL system can be used for stroke risk assessment during routine or clinical trial modes.

Symtosis: A liver ultrasound tissue characterization and risk stratification in optimized deep learning paradigm

Diabetes mellitus is a group of metabolic diseases in which blood sugar levels are too high. About 8.8% of the world was diabetic in 2017. It is projected that this will reach nearly 10% by 2045. The major challenge is that when machine learning-based classifiers are applied to such data sets for risk stratification, leads to lower performance. Thus, our objective is to develop an optimized and robust machine learning (ML) system under the assumption that missing values or outliers if replaced by a median configuration will yield higher risk stratification accuracy. This ML-based risk stratification is designed, optimized and evaluated, where: (i) the features are extracted and optimized from the six feature selection techniques (random forest, logistic regression, mutual information, principal component analysis, analysis of variance, and Fisher discriminant ratio) and combined with ten different types of classifiers (linear discriminant analysis, quadratic discriminant analysis, naïve Bayes, Gaussian process classification, support vector machine, artificial neural network, Adaboost, logistic regression, decision tree, and random forest) under the hypothesis that both missing values and outliers when replaced by computed medians will improve the risk stratification accuracy. Pima Indian diabetic dataset (768 patients: 268 diabetic and 500 controls) was used. Our results demonstrate that on replacing the missing values and outliers by group median and median values, respectively and further using the combination of random forest feature selection and random forest classification technique yields an accuracy, sensitivity, specificity, positive predictive value, negative predictive value and area under the curve as: 92.26%, 95.96%, 79.72%, 91.14%, 91.20%, and 0.93, respectively. This is an improvement of 10% over previously developed techniques published in literature. The system was validated for its stability and reliability. RF-based model showed the best performance when outliers are replaced by median values.

Deep learning fully convolution network for lumen characterization in diabetic patients using carotid ultrasound: a tool for stroke risk

Background and Objective Fatty Liver Disease (FLD) - a disease caused by deposition of fat in liver cells, is predecessor to terminal diseases such as liver cancer. The machine learning (ML) techniques applied for FLD detection and risk stratification using ultrasound (US) have limitations in computing tissue characterization features, thereby limiting the accuracy. Methods Under the class of Symtosis for FLD detection and risk stratification, this study presents a Deep Learning (DL)-based paradigm that computes nearly seven million weights per image when passed through a 22 layered neural network during the cross-validation (training and testing) paradigm. The DL architecture consists of cascaded layers of operations such as: convolution, pooling, rectified linear unit, dropout and a special block called inception model that provides speed and efficiency. All data analysis is performed in optimized tissue region, obtained by removing background information. We benchmark the DL system against the conventional ML protocols: support vector machine (SVM) and extreme learning machine (ELM). Results The liver US data consists of 63 patients (27 normal/36 abnormal). Using the K10 cross-validation protocol (90% training and 10% testing), the detection and risk stratification accuracies are: 82%, 92% and 100% for SVM, ELM and DL systems, respectively. The corresponding area under the curve is: 0.79, 0.92 and 1.0, respectively. We further validate our DL system using two class biometric facial data that yields an accuracy of 99%. Conclusion DL system shows a superior performance for liver detection and risk stratification compared to conventional machine learning systems: SVM and ELM.

Cerebral Small Vessel Disease: A Review Focusing on Pathophysiology, Biomarkers, and Machine Learning Strategies

AbstractManual ultrasound (US)-based methods are adapted for lumen diameter (LD) measurement to estimate the risk of stroke but they are tedious, error prone, and subjective causing variability. We propose an automated deep learning (DL)-based system for lumen detection. The system consists of a combination of two DL systems: encoder and decoder for lumen segmentation. The encoder employs a 13-layer convolution neural network model (CNN) for rich feature extraction. The decoder employs three up-sample layers of fully convolution network (FCN) for lumen segmentation. Three sets of manual tracings were used during the training paradigm leading to the design of three DL systems. Cross-validation protocol was implemented for all three DL systems. Using the polyline distance metric, the precision of merit for three DL systems over 407 US scans was 99.61%, 97.75%, and 99.89%, respectively. The Jaccard index and Dice similarity of DL lumen segmented region against three ground truth (GT) regions were 0.94, 0.94, and 0.93 and 0.97, 0.97, and 0.97, respectively. The corresponding AUC for three DL systems was 0.95, 0.91, and 0.93. The experimental results demonstrated superior performance of proposed deep learning system over conventional methods in literature. Graphical abstractᅟ