Left Ventricle Segmentation Using Scale-Independent Multi-Gate UNET in MRI Images

Introduction The accuracy of echocardiography interpretations is crucial. However, echocardiography has varying degrees of image quality, which can impact diagnostic decision-making. Automated chamber quantification of the left ventricle (LV) and left atrium (LA) has been shown to be feasible using machine learning (ML) systems. However, the effect of image quality on the performance of these ML systems has not been investigated. Purpose To assess the effect of different echocardiographic image quality on the performance of left-sided heart quantification using an ML system. Methods A multi-center study enrolled 240 patients who underwent stress echocardiography across five centers. The acquired rest phase images were anonymized for subsequent analysis. Expert cardiologists reviewed the studies, selecting apical four-chamber (4Ch) and two-chamber (2Ch) view images. The cardiologists identified end-systolic (ES) and end-diastolic (ED) frames. They traced the endocardial borders for the LA in ES and the LV in ED and ES. The same ED and ES frames were then processed by an ML algorithm trained on a separate set of images. Subsequently, an expert cardiologist, blinded to the previous measurements, assessed the image quality of LA and LV. The image quality for each chamber in the ES or ED frame was categorized into one of four grades: poor, suboptimal, fair, and optimal. Agreement analysis between human and AI measurements was performed within each image quality group. Yield was defined as the percentage of studies where the ML system predicted an echocardiographic parameter compared to a cardiologist. Results The yield of study parameters varied significantly with image quality. The lowest yield was 45%, observed for LV end-diastolic volume in the 2Ch view when image quality was poor. In contrast, the yield of 100% was seen for LV end-diastolic volumes in both the 4Ch and 2Ch views for optimal quality images. The correlation between automated and expert cardiologist measurements of left heart volumes varied depending on image quality. For poor or suboptimal images, the Pearson correlation coefficient ranged from 0.66 to 0.91 and improved to 0.85-0.92 in fair and optimal-quality images. Additionally, there was an inverse relation between agreement metrics (bias, limits of agreement, and coefficient of variation) and improvements in image quality. However, this tendency was less pronounced or nonexistent when comparing fair and optimal-quality images. (Table 1). Conclusions Image quality affects the yield of automated analysis and the agreement between cardiologists and automated measurements in echocardiography, highlighting the need for high-quality image acquisition in use with ML systems. Table 1

We report the case of a 44-year-old man referred to our institution for suspicion of arrhythmogenic right ventricular (RV) cardiomyopathy. Both familial and personal medical histories were unremarkable, except the notion of a slight unexplored heart murmur during childhood. He decided to undergo a “40s” medical checkup. On presentation to the general practitioner, he had described a slight fatigue for 8 months without concomitant stress. Physical examination revealed no abnormality. ECG showed a type I atrioventricular block and a complete right bundle-branch block (Figure 1A). Transthoracic echocardiography showed an unusual left ventricle (LV) with a leftward curveted interventricular septum (IVS) and preserved global and segmental LV systolic function, suggestive of high systolic RV pressure or takotsubo cardiomyopathy (Figure 1B and Movie I in the online-only Data Supplement). However further analysis showed normal systolic pulmonary artery pressure and LV filling pressures. The apical 4-chamber view showed an abnormal RV with an aneurysm-like shape of the basal lateral free wall (Figure 1B and Movie II in the online-only Data Supplement) but with normal kinetics. A first cardiac magnetic resonance image showed biventricular abnormalities, particularly focused on RV wall motion, which justified additional morphological investigations. No abnormal pulmonary venous return was noticed. Invasive coronary angiogram was normal. The patient was later referred to our unit by his attending cardiologist and underwent invasive hemodynamics and biventricular contrast angiography. Cardiac index (pulmonary artery thermodilution) and biventricular volumes were normal. Blood gas content analysis disclosed an intracardiac shunt. RV angiography (Figure 2A and 2B) showed a truncated RV apex and the absence of trabeculations in the anteroapical and inferoapical zones, therefore characterized by a smooth …

Automatic segmentation of left ventricle using parallel end–end deep convolutional neural networks framework

Accurate detection of end-systolic (ES) and end-diastolic (ED) frames in an echocardiographic cine series can be difficult but necessary pre-processing step for the development of automatic systems to measure cardiac parameters. The detection task is challenging due to variations in cardiac anatomy and heart rate often associated with pathological conditions. We formulate this problem as a regression problem and propose several deep learning-based architectures that minimize a novel global extrema structured loss function to localize the ED and ES frames. The proposed architectures integrate convolution neural networks (CNNs)-based image feature extraction model and recurrent neural networks (RNNs) to model temporal dependencies between each frame in a sequence. We explore two CNN architectures: DenseNet and ResNet, and four RNN architectures: long short-term memory, bi-directional LSTM, gated recurrent unit (GRU), and Bi-GRU, and compare the performance of these models. The optimal deep learning model consists of a DenseNet and GRU trained with the proposed loss function. On average, we achieved 0.20 and 1.43 frame mismatch for the ED and ES frames, respectively, which are within reported inter-observer variability for the manual detection of these frames.

Clinical indications of heart disease are shown from left ventricle (LV) or right ventricle (RV) volume measurements of cardiac MRI images. LV and RV segmentation of cardiac MRI images can detect and measure image volume. Public dataset MICCAI, ACDC, Kaggle, and SCD provide data on MRI images of cardiac that have been widely used by researchers. The deep learning method approach can optimally solve problems in analyzing heart disease from cardiac MRI images. The aim of this paper is to determine the availability of public datasets that are appropriate for the research objectives. It can support the optimization of the segmentation method for LV and RV images of cardiac as the contribution of this paper. The results of the study are that the public dataset (MICCAI, ACDC, Kaggle, and SCD) provides sufficient data for the identification, classification, and measurement of LV and RV volumes. Furthermore, a deep learning approach with convolutional neural networks can detect and classify heart diseases with high accuracy.

Semi-supervised learning improves the performance of cardiac event detection in echocardiography

The authors propose a novel spatiotemporal constraint based on shape and appearance and combine it with a level‐set deformable model for left ventricle (LV) segmentation in four‐dimensional gated cardiac SPECT, particularly in the presence of perfusion defects. The model incorporates appearance and shape information into a ‘soft‐to‐hard’ probabilistic constraint, and utilises spatiotemporal regularisation via a maximum a posteriori framework. This constraint force allows more flexibility than the rigid forces of shape constraint‐only schemes, as well as other state of the art joint shape and appearance constraints. The combined model can hypothesise defective LV borders based on prior knowledge. The authors present comparative results to illustrate the improvement gain. A brief defect detection example is finally presented as an application of the proposed method.

The domain of deep learning stimulates medical image analysis, which is a catalyst of scientific research and an essential element of healthcare. Since semantic segmentation techniques empower image processing and quantitative determination in various applications, designing a dedicated solution is challenging and heavily reliant on input data characteristics and hardware constraints. Cardiac magnetic resonance imaging segmentation provides three essential heart structures, left ventricle (LV) cavity, myocardium (MYO), and right ventricle (RV) cavity. In clinical applications, manual contouring is frequently utilized to perform semantic segmentation. A fully automated cardiac magnetic resonance image (CMRI) segmentation technique is becoming more desirable as deep learning-based frameworks advance. Motivated by the power of U-Net, residual network, and deep supervision, this paper proposes a Deep Residual U-Net to achieve better LV, MYO, and RV segmentation in short-axis CMRI. The model is formed with residual connections and has a similar layout to U-Net. It provides three advantages: First, residual connections make deep network training easier. Second, the network's rich skip connections enable feature propagation, allowing for network creation with lower complexity but more remarkable performance. Third, Deep supervision facilitates the loss computation at every feature dimension except at least two, empowering gradients to be implanted more deeply in the network and improving each layer's training in Deep Residual U-Net. Automated Cardiac Diagnosis Challenge (ACDC) 2017 dataset has been used to assess the segmentation efficiency of proposed model. The proposed approach significantly outperformed all other methods, evidencing its supremacy over recent state-of-the-art U-Net-based techniques.

Generation of parametric images of myocardial blood flow (MBF) from dynamic PET requires a reproducible and reliable method to obtain left and right ventricular time-activity curves. We propose the use of Independent Component Analysis (ICA) on PET temporal information to segment the image data into three categories: right ventricle (RV), left ventricle (LV) and myocardium. ICA is an established signal processing technique that assumes that the dynamic signal in each voxel is a linear mixture of signals from statistically independent sources. The mixing process is due to cardiac motion and partial volume effects. Due to its ability to find the intensity of the underlying sources, ICA generates a 3D map of the mixing coefficients that allows the segmentation of structures. Dynamic cardiac PET data were acquired with a Siemens Biograph mCT PET/CT scanner using Rb-82 and [15O]water in nonhuman primate and pigs with injections of different activities. The Dice similarity coefficient (DSC) was used to compare the segmentations between scans of different species with different tracers. The LV and RV regions were used to calculate input functions for kinetic analysis and MBF estimation. Pairwise comparisons of the segmentations between different [15O]water PET scans gave DSC of 0.91 ± 0.09 and 0.92±0.08 for the RV and LV segmentations, respectively. For Rb-82 PET scans, the DSC values were of 0.89 ± 0.09 and 0.85±0.10 for RV and LV. We used contrast CT as a gold standard for the LV and RV segmentations. Pairwise comparisons of the segmentations between different contrast CT and Rb-82 PET scans gave DSC of 0.86 ± 0.09 and 0.85±0.09 for the RV and LV, respectively. Our work shows the reliability and reproducibility of the results of the segmentation algorithm. High quality parametric images of MBF and the kinetic parameters were obtained. These results suggest that ICA may be useful to extract time-activity curves for kinetic MBF analysis. The effects of partial volume on input TAC determination were also evaluated.

In this work we automatically segment the left ventricle (LV) in cardiac MR images in the end-diastole (ED) and end-systole (ES) phases using a novel approach that combines statistical and deterministic deformable models. A 3D Active Appearance Model (AAM) is used to segment the ED phase. The AAM texture model is trained on radial samples from gradient magnitude images to make the fitting process faster and more discriminative. A trained ED-to-ES shape correspondence model is used to map a given ED shape to an ES shape. Once the AAM model converges to a shape in ED, the correspondence model is used to get an approximate ES shape. We segment the LV in the ES phase by first fitting a deformable superquadric to the AAM converged shape (in ED) using data range forces and then tracking the LV using image and data range forces (for the ES shape obtained from correspondence model). We test our approach by performing leave-one-out training on a 35 patient datasets. The data comprises 19 normal patients and 16 patients having heart abnormalities (cardiomyopathy and myocardial infarction). The composition makes it a challenging data collection with significant shape variation. The performance of our method is evaluated by measuring the mismatch between automatically segmented and expert delineated contours using the Mean Perpendicular Distance (MPD) and Dice metrics. The average MPD is 2.6mm for ED and 3.7mm for ES (error mostly towards the apex and base). The average Dice is 0.9 for ED and 0.8 for ES. These results show good potential for clinical use.

IntroductionAlthough global deformation parameters have been increasingly used for myocardial function analysis, there are sparse data concerning segmental deformation of the left ventricle (LV). Moreover, some studies suggest heterogeneity of strain among LV segments, which may be especially significant during stress echocardiography. We assessed quantitatively regional LV function in the setting of dobutamine stress echocardiography (DSE), to examine differences of longitudinal strain between basal, mid and apical LV segments and to compare variability of regional deformation between rest and the peak stage of DSE.Material and methodsAmong 250 patients examined by DSE applied for diagnosis of ischemia, a subset of 111 patients without significant coronary stenoses in angiography was selected (68 females, mean age: 60 ±10 years). Systolic longitudinal strain (SLS) in individual LV segments at baseline and the peak stage of DSE was analyzed with speckle tracking echocardiography.ResultsInhomogeneity of SLS among the LV segments (p < 0.001) was observed at baseline and the peak stage. Dispersion indices were higher at the peak stage of DSE than at baseline (p < 0.001), and the lowest heterogeneity was observed among mid segments. The analysis of changes in SLS during DSE showed SLS reduction in basal and mid-ventricular segments and an increase in apical segments.ConclusionsSignificant heterogeneity of strain and the opposite direction of the longitudinal strain changes during DSE between apical and basal LV segments were observed. This variability among non-ischemic LV segments ought to be considered in quantification of LV function during DSE.

Magnetic Resonance Imaging (MRI) has evolved as a clinical standard-of-care imaging modality for cardiac morphology, function assessment, and guidance of cardiac interventions. All these applications rely on accurate extraction of the myocardial tissue and blood pool from the imaging data. Here we propose a framework for left ventricle (LV) segmentation from cardiac cine MRI. First, we segment the LV blood pool using iterative graph cuts, and subsequently use this information to segment the myocardium. We formulate the segmentation procedure as an energy minimization problem in a graph subject to the shape prior obtained by label propagation from an average atlas using affine registration. The proposed framework has been validated on 30 patient cardiac cine MRI datasets available through the STACOM LV segmentation challenge and yielded fast, robust, and accurate segmentation results.

Most work on left ventricle (LV) ultrasound image segmentation using deep learning has focused on single-frame segmentation of end-diastole (ED) and end-systole (ES) frames. Using these neural network models on the entire cardiac cycle often results in segmentation flickering and sudden large segmentation errors. Neural networks that perform some form of temporal reasoning is needed to solve these issues. In this work, we have investigated the use of neural networks with convolutional long short-term memory (ConvLSTM) layers for real-time temporal coherent LV segmentation. A comparison on a dataset of 174 apical 4-, 3- and 2-chamber ultrasound recordings indicated that increasing the number of frames annotated from the cardiac cycle improves temporal segmentation, while using weighted moving average post processing can reduce segmentation flickering, and using ConvLSTM layers reduces large temporal errors considerably. The runtime of the ConvLSTM segmentation network was 13 ms when used in a real-time application for automatic ejection fraction.

Introduction: The extent that left atrium (LA) functional remodeling is impacted by left ventricle (LV) systolic function or clinical history in patients with chronic heart failure (HF) is unresolved. Cardiac imaging and plasma biomarkers may provide valuable information in this regard. Hypothesis: Elastic and active contributions to LV stroke volume (LV SV) are variable, independent of LV systolic function and reflected in natriuretic peptide measurements. Methods: We recruited 27 controls without overt cardiovascular disease and 68 patients with moderate HF, including preserved and reduced LV ejection fraction (LVEF). Cardiac MRI was used to evaluate LV volumes at end diastole and end systole, and LA volumes at end diastole, diastasis and end systole to determine passive and active volume changes. Plasma mid-region pro-atrial and N-terminal pro-brain natriuretic peptides (MR-proANP and NT-proBNP, respectively) were measured. Results: Elastic LA contribution to LV SV was significantly lower in HF (median, IQR: 14%, 8-23%) than controls (24%, 17-30%; p=0.001), but active contribution was not (23%, 0-34% vs 24%, 21-31%; p=0.163). In patients with HF, active LA contribution to LV SV was negatively associated with history of hypertension (β=-0.41, p=0.007) and positively associated with aging (β=0.40, p=0.008) in a multiple linear regression model. Seven of nine subjects with HF and a history of atrial fibrillation (AF), but in sinus rhythm (SR) on the day of study, had active LA contributions to LV SV &gt;15%. Increased LA volume at end diastole (β=0.51, p=0.001) and LVEF (β=-0.33, p=0.008) were independently associated with NT-proBNP (β=0.51, p=0.001) in a model that included age and history of AF. Similarly, LA volume at end diastole (β=0.48, p=0.001) was independently associated with MR-proANP in a model that included age and history of AF. Notably, natriuretic propeptides were not associated with elastic or active LA contributions to LV SV. Conclusions: Heart failure is characterized by preserved LA active contribution to LV SV, regardless of LVEF, which increases with age, and is modifiable by hypertension and AF. Imaging-based assessment can elucidate the contribution of LA function in HF and may provide insight into the impact of AF in patients with HF.

The left ventricle segmentation (LVS) is of great important for the evaluation of cardiac function. This study aimed to establish new segmentation algorithms that can enhance the accuracy and robustness of automatic LVS on magnetic resonance images. The datasets involved 45 subjects, including 12 heart failure patients with ischemia, 12 heart failure patients without ischemia, 12 hypertrophy patients and 9 normal individuals. The experiments consisted of three important steps. At first, deep learning was employed for the coarse LVS on myocardial images. Next, a double snake model was applied to assess the endo- and epi-cardial boundaries. Finally, the optimal epicardial boundary was obtained by adopting radial region growing method. Additionally, the performance of the developed LVS method was evaluated by the previously established software. Furthermore, the developed LVS method was validated by applying the datasets of 45 subjects. The results showed that the good contours, overlapping dice metric and average perpendicular distance of both epi- and endo-cardial contours were approximately 97%, 0.97 and 1.8 mm respectively. The regression coefficient and coefficient of determination between the proposed method and clinical experts were 0.96 and 1.039, respectively for ejection fraction, while 0.92 and 0.994 for left ventricle mass. These findings reveal that the developed method can enhance the accuracy and robustness of LVS. This novel LVS approach exhibits outstanding performance and possesses promising potential to increase the reliability of computer-aided imaging detection system for cardiovascular disease.