BEAS-Net: a Shape-Prior-Based Deep Convolutional Neural Network for Robust Left Ventricular Segmentation in 2D Echocardiography.

Left ventricle (LV) segmentation of 2D echocardiography images is an essential step in the analysis of cardiac morphology and function and - more generally - diagnosis of cardiovascular diseases. Several deep learning (DL) algorithms have recently been proposed for the automatic segmentation of the LV, showing significant performance improvement over the traditional segmentation algorithms. However, unlike the traditional methods, prior information about the segmentation problem, e.g. anatomical shape information, is not usually incorporated for training the DL algorithms. This can degrade the generalization performance of the DL models on unseen images if their characteristics are somewhat different from those of the training images, e.g. low-quality testing images. In this study, a new shape-constrained deep convolutional neural network (CNN) - called BEAS-Net - is introduced for automatic LV segmentation. The BEAS-Net learns how to associate the image features, encoded by its convolutional layers, with anatomical shape-prior information derived by the B-spline explicit active surface (BEAS) algorithm to generate physiologically meaningful segmentation contours when dealing with artifactual or low-quality images. The performance of the proposed network was evaluated using three different in-vivo datasets and was compared a deep segmentation algorithm based on the U-Net model. Both networks yielded comparable results when tested on images of acceptable quality, but the BEAS-Net outperformed the benchmark DL model on artifactual and low-quality images.

A machine learning framework for the evaluation of myocardial rotation in patients with noncompaction cardiomyopathy.

AIMS: Noncompaction cardiomyopathy (NCC) is considered a genetic cardiomyopathy with unknown pathophysiological mechanisms. We propose to evaluate echocardiographic predictors for rigid body rotation (RBR) in NCC using a machine learning (ML) based model. METHODS AND RESULTS: Forty-nine outpatients with NCC diagnosis by echocardiography and magnetic resonance imaging (21 men, 42.8±14.8 years) were included. A comprehensive echocardiogram was performed. The layer-specific strain was analyzed from the apical two-, three, four-chamber views, short axis, and focused right ventricle views using 2D echocardiography (2DE) software. RBR was present in 44.9% of patients, and this group presented increased LV mass indexed (118±43.4 vs. 94.1±27.1g/m2, P = 0.034), LV end-diastolic and end-systolic volumes (P< 0.001), E/e' (12.2±8.68 vs. 7.69±3.13, P = 0.034), and decreased LV ejection fraction (40.7±8.71 vs. 58.9±8.76%, P < 0.001) when compared to patients without RBR. Also, patients with RBR presented a significant decrease of global longitudinal, radial, and circumferential strain. When ML model based on a random forest algorithm and a neural network model was applied, it found that twist, NC/C, torsion, LV ejection fraction, and diastolic dysfunction are the strongest predictors to RBR with accuracy, sensitivity, specificity, area under the curve of 0.93, 0.99, 0.80, and 0.88, respectively. CONCLUSION: In this study, a random forest algorithm was capable of selecting the best echocardiographic predictors to RBR pattern in NCC patients, which was consistent with worse systolic, diastolic, and myocardium deformation indices. Prospective studies are warranted to evaluate the role of this tool for NCC risk stratification.

Non-rigid image registration using a modified fuzzy feature-based inference system for 3D cardiac motion estimation.

BACKGROUND AND OBJECTIVE: Non-rigid image registration is a well-established method for estimating cardiac motion on 3D echocardiographic images. However, such images have relatively poor spatio-temporal resolution making registration challenging. Some of the main challenges are extracting features relevant to the registration problem and defining a suitable geometrical transformation to be applied. The latter can be tackled using a fuzzy inference system considering its potential in transformation modeling. From this point of view, feature-based image registration can be considered an identification problem in which the transformation parameters are computed through an optimization process. This study, thus, aims to estimate cardiac motion on 3D echocardiographic images based on feature-based non-rigid image registration through sets of modified fuzzy rules. METHODS: The 3D volume features were extracted with the popular scale-invariant feature transform (SIFT) descriptors in 3D space. Sets of fuzzy rules were generated according to the extracted features to register every two consecutive frames. Finally, some supplementary rules modified the registration rule for estimating cardiac motion. RESULTS: Applying the fuzzy feature-based inference system on the STRAUS synthetic database showed the proposed method to be competitive with other well-established registration algorithms in terms of tracking error and accuracy of strain estimates. The proposed algorithm yielded a tracking error of 1 mm and a relative circumferential strain error of 0.82±4.69%. In addition, the potential of the proposed algorithm for clinical applications was confirmed by evaluating its performance on an in-vivo database called CETUS. CONCLUSION: This paper proposes a novel registration method based on fuzzy logic which was shown to enable tracking complex cardiac deformations in 3D echocardiographic images with high accuracy.

Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME): A Checklist: Reviewed by the American College of Cardiology Healthcare Innovation Council.

Machine learning (ML) has been increasingly used within cardiology, particularly in the domain of cardiovascular imaging. Due to the inherent complexity and flexibility of ML algorithms, inconsistencies in the model performance and interpretation may occur. Several review articles have been recently published that introduce the fundamental principles and clinical application of ML for cardiologists. This paper builds on these introductory principles and outlines a more comprehensive list of crucial responsibilities that need to be completed when developing ML models. This paper aims to serve as a scientific foundation to aid investigators, data scientists, authors, editors, and reviewers involved in machine learning research with the intent of uniform reporting of ML investigations. An independent multidisciplinary panel of ML experts, clinicians, and statisticians worked together to review the theoretical rationale underlying 7 sets of requirements that may reduce algorithmic errors and biases. Finally, the paper summarizes a list of reporting items as an itemized checklist that highlights steps for ensuring correct application of ML models and the consistent reporting of model specifications and results. It is expected that the rapid pace of research and development and the increased availability of real-world evidence may require periodic updates to the checklist.

Biventricular imaging markers to predict outcomes in non-compaction cardiomyopathy: a machine learning study.

AIMS: Left ventricular non-compaction cardiomyopathy (LVNC) is a genetic heart disease, with heart failure, arrhythmias, and embolic events as main clinical manifestations. The goal of this study was to analyse a large set of echocardiographic (echo) and cardiac magnetic resonance imaging (CMRI) parameters using machine learning (ML) techniques to find imaging predictors of clinical outcomes in a long-term follow-up of LVNC patients. METHODS AND RESULTS: Patients with echo and/or CMRI criteria of LVNC, followed from January 2011 to December 2017 in the heart failure section of a tertiary referral cardiologic hospital, were enrolled in a retrospective study. Two-dimensional colour Doppler echocardiography and subsequent CMRI were carried out. Twenty-four hour Holter monitoring was also performed in all patients. Death, cardiac transplantation, heart failure hospitalization, aborted sudden cardiac death, complex ventricular arrhythmias (sustained and non-sustained ventricular tachycardia), and embolisms (i.e. stroke, pulmonary thromboembolism and/or peripheral arterial embolism) were registered and were referred to as major adverse cardiovascular events (MACEs) in this study. Recruited for the study were 108 LVNC patients, aged 38.3 ± 15.5 years, 48.1% men, diagnosed by echo and CMRI criteria. They were followed for 5.8 ± 3.9 years, and MACEs were registered. CMRI and echo parameters were analysed via a supervised ML methodology. Forty-seven (43.5%) patients had at least one MACE. The best performance of imaging variables was achieved by combining four parameters: left ventricular (LV) ejection fraction (by CMRI), right ventricular (RV) end-systolic volume (by CMRI), RV systolic dysfunction (by echo), and RV lower diameter (by CMRI) with accuracy, sensitivity, and specificity rates of 75.5%, 77%, 75%, respectively. CONCLUSIONS: Our findings show the importance of biventricular assessment to detect the severity of this cardiomyopathy and to plan for early clinical intervention. In addition, this study shows that even patients with normal LV function and negative late gadolinium enhancement had MACE. ML is a promising tool for analysing a large set of parameters to stratify and predict prognosis in LVNC patients.

Area of the pressure-strain loop during ejection as non-invasive index of left ventricular performance: a population study.

BACKGROUND: Previous studies highlighted the usefulness of integrating left ventricular (LV) deformation (strain) and hemodynamic parameters to quantify LV performance. In a population sample, we investigated the anthropometric and clinical determinants of a novel non-invasive index of LV systolic performance derived from simultaneous registration of LV strain and brachial pressure waveforms. METHODS: Three hundred fifty-six randomly recruited subjects (44.7% women; mean age, 53.9 years; 47.5% hypertensive) underwent echocardiographic and arterial data acquisition. We constructed pressure-strain loops from simultaneously recorded two-dimensional LV strain curves and brachial pressure waveforms obtained by finger applanation tonometry. We defined the area of this pressure-strain loop during ejection as LV ejection work density (EWD). We reported effect sizes as EWD changes associated with a 1-SD increase in covariables. RESULTS: In multivariable-adjusted analyses, higher EWD was associated with age, female sex and presence of hypertension (P ≤ 0.0084). In both men and women, EWD increased independently with augmentation pressure (effect size: + 59.1 Pa), central pulse pressure (+ 65.7 Pa) and pulse wave velocity (+ 44.8 Pa; P ≤ 0.0006). In men, EWD decreased with relative wall thickness (- 29.9 Pa) and increased with LV ejection fraction (+ 23.9 Pa; P ≤ 0.040). In women, EWD increased with left atrial (+ 76.2 Pa) and LV end-diastolic (+ 43.8 Pa) volume indexes and with E/e' ratio (+ 51.1 Pa; P ≤ 0.026). CONCLUSION: Older age, female sex and hypertension were associated with higher EWD. Integration of the LV pressure-strain loop during ejection might be a useful tool to non-invasively evaluate sex-specific and interdependent effects of preload and afterload on LV myocardial performance.

Diagnosis of Heart Failure With Preserved Ejection Fraction: Machine Learning of Spatiotemporal Variations in Left Ventricular Deformation.

BACKGROUND: Stress testing helps diagnose heart failure with preserved ejection fraction (HFpEF), but there are no established criteria for quantifying left ventricular (LV) functional reserve. The aim of this study was to investigate whether comprehensive analysis of the timing and amplitude of LV long-axis myocardial motion and deformation throughout the cardiac cycle during rest and stress can provide more informative criteria than standard measurements. METHODS: Velocity, strain, and strain rate traces were measured from all 18 LV segments by echocardiographic myocardial velocity imaging at rest and during semisupine bicycle exercise in 100 subjects aged 69 ± 7 years, including patients with HFpEF and healthy, hypertensive, and breathless control subjects. A machine-learning algorithm, composed of an unsupervised statistical method and a supervised classifier, was used to model spatiotemporal patterns of the traces and compare the predicted labels with the clinical diagnoses. RESULTS: The learned strain rate parameters gave the highest accuracy for allocating subjects into the four groups (overall, 57%; for patients with HFpEF, 81%), and into two classes (asymptomatic vs symptomatic; area under the curve, 0.89; accuracy, 85%; sensitivity, 86%; specificity, 82%). Machine learning of strain rate, compared with standard measurements, gave the greatest improvement in accuracy for the two-class task (+23%, P < .0001), compared with +11% (P < .0001) using velocity and +4% (P < .05) using strain. Strain rate was also best at predicting 6-min walk distance as an independent reference criterion. CONCLUSIONS: Machine learning of spatiotemporal variations of LV strain rate during rest and exercise could be used to identify patients with HFpEF and to provide an objective basis for diagnostic classification.

Statistical shape modeling of the left ventricle: myocardial infarct classification challenge.

Statistical shape modeling is a powerful tool for visualizing and quantifying geometric and functional patterns of the heart. After myocardial infarction (MI), the left ventricle typically remodels in response to physiological challenges. Several methods have been proposed in the literature to describe statistical shape changes. Which method best characterizes left ventricular remodeling after MI is an open research question. A better descriptor of remodeling is expected to provide a more accurate evaluation of disease status in MI patients. We therefore designed a challenge to test shape characterization in MI given a set of three-dimensional left ventricular surface points. The training set comprised 100 MI patients, and 100 asymptomatic volunteers (AV). The challenge was initiated in 2015 at the Statistical Atlases and Computational Models of the Heart workshop, in conjunction with the MICCAI conference. The training set with labels was provided to participants, who were asked to submit the likelihood of MI from a different (validation) set of 200 cases (100 AV and 100 MI). Sensitivity, specificity, accuracy and area under the receiver operating characteristic curve were used as the outcome measures. The goals of this challenge were to (1) establish a common dataset for evaluating statistical shape modeling algorithms in MI, and (2) test whether statistical shape modeling provides additional information characterizing MI patients over standard clinical measures. Eleven groups with a wide variety of classification and feature extraction approaches participated in this challenge. All methods achieved excellent classification results with accuracy ranges from 0.83 to 0.98. The areas under the receiver operating characteristic curves were all above 0.90. Four methods showed significantly higher performance than standard clinical measures. The dataset and software for evaluation are available from the Cardiac Atlas Project website1.

Machine learning of the spatio-temporal characteristics of echocardiographic deformation curves for infarct classification.

The aim of this study was to analyze the whole temporal profiles of the segmental deformation curves of the left ventricle (LV) and describe their interrelations to obtain more detailed information concerning global LV function in order to be able to identify abnormal changes in LV mechanics. The temporal characteristics of the segmental LV deformation curves were compactly described using an efficient decomposition into major patterns of variation through a statistical method, called Principal Component Analysis (PCA). In order to describe the spatial relations between the segmental traces, the PCA-derived temporal features of all LV segments were concatenated. The obtained set of features was then used to build an automatic classification system. The proposed methodology was applied to a group of 60 MRI-delayed enhancement confirmed infarct patients and 60 controls in order to detect myocardial infarction. An average classification accuracy of 87% with corresponding sensitivity and specificity rates of 89% and 85%, respectively was obtained by the proposed methodology applied on the strain rate curves. This classification performance was better than that obtained with the same methodology applied on the strain curves, reading of two expert cardiologists as well as comparative classification systems using only the spatial distribution of the end-systolic strain and peak-systolic strain rate values. This study shows the potential of machine learning in the field of cardiac deformation imaging where an efficient representation of the spatio-temporal characteristics of the segmental deformation curves allowed automatic classification of infarcted from control hearts with high accuracy.