Skip to main content

About this book

This book constitutes the thoroughly refereed post-workshop proceedings of the 9th International Workshop on Statistical Atlases and Computational Models of the Heart: Atrial Segmentation and LV Quantification Challenges, STACOM 2018, held in conjunction with MICCAI 2018, in Granada, Spain, in September 2018.
The 52 revised full workshop papers were carefully reviewed and selected from 60 submissions. The topics of the workshop included: cardiac imaging and image processing, machine learning applied to cardiac imaging and image analysis, atlas construction, statistical modelling of cardiac function across different patient populations, cardiac computational physiology, model customization, atlas based functional analysis, ontological schemata for data and results, integrated functional and structural analyses, as well as the pre-clinical and clinical applicability of these methods.

Table of Contents


Regular Papers


Estimating Sheets in the Heart Wall

Models of sheets in the heart wall play an important role in the visualization of myofiber geometry, in modelling mechanics and in cardiac electrophysiology. For example, the assumption of distinct speeds of propagation in the directions of myofibers, the sheets in which they lie, and the direction across them, can predict the arrival time of the conduction wave that triggers myocyte contraction. Almost all current analyses based on DTI data use the third eigenvector of the diffusion tensor as an estimate of the local sheet normal. This construction suffers from the limitation that the second and third eigenvector directions can be ambiguous since they are associated with eigenvalues that are quite similar. Here we present and evaluate an alternate method to estimate sheets, which uses only the principal eigenvector. We find the best local direction perpendicular to the principal eigenvector to span a sheet, using the Lie bracket and the minimization of an appropriately constructed energy function. We test our method on a dataset of 8 ex vivo rat and 8 ex vivo canine cardiac diffusion tensor images. Qualitatively the recovered sheets are more consistent with the geometry of myofibers than those obtained using all three eigenvectors, particularly when they curve or fan. Quantitatively the recovered sheet normals also give a low value of holonomicity, a measure of the degree to which they are orthogonal to a family of surfaces. Our novel fitting approach could thus impact cardiac mechanical and electrophysiological analyses which are based on DTI data.

Tabish A. Syed, Babak Samari, Kaleem Siddiqi

Automated Motion Correction and 3D Vessel Centerlines Reconstruction from Non-simultaneous Angiographic Projections

Automated estimation of 3D centerlines is necessary to transform angiographic projections into accurate 3D reconstructions. Although several methods exist for 3D centerline reconstruction, most of them are sensitive to the motion in coronary arteries when images are acquired by a single-plane rotational X-ray system. The objective of the proposed method is to rectify the motion-related deformations in coronary vessels from 2D projections and subsequently achieve an optimal 3D centerline reconstruction. Rigid motion in arteries is removed by estimating the optimal rigid transformation from all projection planes. The remaining non-rigid motion at end-diastole is modelled by a radial basis function based warping of 2D centerlines. Point correspondences are then generated from all projection planes by least squares matching. The final 3D centerlines are obtained by 3D non-uniform rational basis splines fitting over generated point correspondences. Experimental analysis over 20 coronary vessel trees (12 right coronary artery: RCA and 8 left coronary artery: LCA) demonstrates that the rigid transformation is able to reduce the coronary vessel movements to 0.72 mm average, while the final 3D centerline reconstruction achieves an average rms error of 0.31 mm, when backprojected on angiographic planes.

Abhirup Banerjee, Rajesh K. Kharbanda, Robin P. Choudhury, Vicente Grau

Left Ventricle Segmentation and Quantification from Cardiac Cine MR Images via Multi-task Learning

Segmentation of the left ventricle and quantification of various cardiac contractile functions is crucial for the timely diagnosis and treatment of cardiovascular diseases. Traditionally, the two tasks have been tackled independently. Here we propose a convolutional neural network based multi-task learning approach to perform both tasks simultaneously, such that, the network learns better representation of the data with improved generalization performance. Probabilistic formulation of the problem enables learning the task uncertainties during the training, which are used to automatically compute the weights for the tasks. We performed a five fold cross-validation of the myocardium segmentation obtained from the proposed multi-task network on 97 patient 4-dimensional cardiac cine-MRI datasets available through the STACOM LV segmentation challenge against the provided gold-standard myocardium segmentation, obtaining a Dice overlap of $$0.849 \pm 0.036$$ and mean surface distance of $$0.274 \pm 0.083$$ mm, while simultaneously estimating the myocardial area with mean absolute difference error of $$205\pm 198\,\mathrm{mm}^2$$ .

Shusil Dangi, Ziv Yaniv, Cristian A. Linte

Statistical Shape Clustering of Left Atrial Appendages

Fifteen percent of all strokes are caused by emboli formed in the left atrium (LA) in case of atrial fibrillation (AF). The most common site of thrombus formation is inside the left atrial appendage (LAA). The LAA is accounting for 70% to 90% of the thrombi formed in the LA in patients with non-valvular AF. Studies have shown there is a correlation between the LAA morphology and risk of ischemic stroke; Chicken Wing and Cauliflower LAA shapes are associated with lower and higher risk, respectively. These two LAA shape categories come from a popular classification in the medical domain, but it is subjective and based on qualitative shape parameters. In this paper, we describe a full framework for shape analysis and clustering of the LAA. Initially, we build a point distribution model to quantitatively describe the LAA shape variation based on 103 LAA surfaces segmented and reconstructed from multidetector computed tomography volumes. We are successfully able to determine point correspondence between LAA surfaces, by non-rigid volumetric registration of signed distance fields. To validate if LAA shapes are clustered, we employ an unsupervised clustering on the shape models parameters to estimate the natural number of clusters in our training set, where the number of shape clusters is estimated by validating the test log-likelihood of several Gaussian mixture models using two level cross-validation. We found that the LAAs surfaces basically formed two shape clusters broadly corresponding to the Chicken wing and non-Chicken Wing morphologies, which fits well with clinical knowledge.

Jakob M. Slipsager, Kristine A. Juhl, Per E. Sigvardsen, Klaus F. Kofoed, Ole De Backer, Andy L. Olivares, Oscar Camara, Rasmus R. Paulsen

Deep Learning Segmentation of the Left Ventricle in Structural CMR: Towards a Fully Automatic Multi-scan Analysis

In the past three years, with the novel use of artificial intelligence for medical image analysis, many authors have focused their efforts on defining automatically the ventricular contours in cardiac Cine MRI. The accuracy reached by deep learning methods is now high enough for routine clinical use. However, integration of other cardiac MR sequences that are routinely acquired along with the functional Cine MR has not been investigated. Namely, T1 maps are static T1-based images that encode in each pixel the T1 relaxation time of the tissue, enabling the definition of local and diffuse fibrosis; T2 maps are static T2-based images that highlight excess water (edema) within the muscle; Late Gadolinium Enhancement (LGE) images are acquired 10 min after injection of a contrast agent that will linger in infarct areas. These sequences are acquired in short-axis plane similar to the 2D Cine MRI, and therefore contain similar anatomical features. In this paper we focus on segmenting the left ventricle in these structural images for further physiological quantification. We first evaluate the use of transfer learning from a model trained on Cine data to analyze these short-axis structural sequences. We also develop an automatic slice selection method to avoid over-segmentation which can be critical in scar/fibrosis/edema delineation. We report good accuracy with dice scores around 0.9 for T1 and T2 maps and correlation of the physiological parameters above 0.9 using only 40 scans and executed in less than 15 s on CPU.

Hakim Fadil, John J. Totman, Stephanie Marchesseau

Cine and Multicontrast Late Enhanced MRI Registration for 3D Heart Model Construction

Cardiac MR imaging using multicontrast late enhancement (MCLE) acquisition provides a way to identify myocardium infarct scar and arrhythmia foci in the peri-infarct. In image-guided RF ablations of ventricular arrhythmia and computational modeling of cardiac function, construction of a 3D heart model is required but this is hampered by the challenges in myocardium segmentation and slice misalignment in MCLE images. Here we developed an approach for cine and MCLE registration, and MCLE scar-cine myocardium label fusion to build high-fidelity 3D heart models. MCLE-cine image alignment was initialized using a block-matching-based rigid registration approach followed by a deformable registration refinement step. The deformable registration approach employed a self similarity context descriptor for image similarity measurements, optical flow as a transformation model and convex optimization to derive the optimal solution. We applied the developed approach to a preclinical dataset of 10 pigs with myocardium infarction and evaluated the registration accuracy by comparing cine and MCLE myocardium masks using Dice-similarity-coefficient (DSC) and average symmetric surface distance (ASSD). For 10 pigs, we achieved a mean DSC of $$80.4\pm 7.8$$ % and ASSD of $$1.28\pm 0.47$$ mm for myocardium with a mean runtime of 1.5 min for each dataset. These results suggest that the developed approach provide the registration accuracy and computational efficiency that may be suitable for clinical applications of cardiac MRI that involve a cine and MCLE MRI registration component.

Fumin Guo, Mengyuan Li, Matthew Ng, Graham Wright, Mihaela Pop

Joint Analysis of Personalized In-Silico Haemodynamics and Shape Descriptors of the Left Atrial Appendage

The left atrial appendage (LAA) is a complex and heterogeneous bulge structure of the left atrium (LA). It is known that, in atrial fibrillation (AF) patients, around 70% to 90% of the thrombi are formed there. However, the exact mechanism of the process of thrombus formation and the role of the LAA in that process are not fully understood yet. The main goal of this work is to perform patient-specific haemodynamics simulations of the LA and LAA and jointly analyse the resulting blood flow parameters with morphological descriptors of these structures in relation with the risk of thrombus formation. Some LAA morphological descriptors such as ostium characteristics and pulmonary configuration were found to influence LAA blood flow patterns. These findings improve our knowledge on the required conditions for thrombus formation in the LAA.

Jordi Mill, Andy L. Olivares, Etelvino Silva, Ibai Genua, Alvaro Fernandez, Ainhoa Aguado, Marta Nuñez-Garcia, Tom de Potter, Xavier Freixa, Oscar Camara

How Accurately Does Transesophageal Echocardiography Identify the Mitral Valve?

Mitral Valve Disease (MVD) describes a variety of pathologies that cause regurgitation of blood during the systolic phase of the cardiac cycle. Decisions in valvular disease management rely heavily on non-invasive imaging. Transesophageal echocardiography (TEE) is widely recognized as the key evaluation technique where backflow of high velocity blood can be visualized under Doppler. However, the heavy reliance on TEE segmentation for diagnosis and modelling necessitates an evaluation of the accuracy of this oft-used mitral valve imaging modality. In this pilot study, we acquire simultaneous CT and TEE images of both a silicone mitral valve phantom and an iodine-stained bovine mitral valve. We propose a pipeline to use CT as ground truth to study the relationship between TEE intensities and the underlying valve morphology. Preliminary results demonstrate that even with an optimized threshold selection based solely on TEE pixel intensities, only 40% of pixels are correctly classified as part of the valve. In addition, we have shown that emphasizing the center line rather than the boundaries of the high intensity regions in TEE provides a better representation and segmentation of the valve morphology. The root mean squared distance between the TEE and CT ground truth is 1.80 mm with intensity-based segmentation and improves to 0.81 mm when comparing the center line extracted from the segmented volumes.

Claire Vannelli, Wenyao Xia, John Moore, Terry Peters

Stochastic Model-Based Left Ventricle Segmentation in 3D Echocardiography Using Fractional Brownian Motion

A novel approach for fully-automated segmentation of the left ventricle (LV) endocardial and epicardial contours is presented. This is mainly based on the natural physical characteristics of the LV shape structure. Both sides of the LV boundaries exhibit natural elliptical curvatures by having details on various scales, i.e. exhibiting fractal-like characteristics. The fractional Brownian motion (fBm), which is a non-stationary stochastic process, integrates well with the stochastic nature of ultrasound echoes. It has the advantage of representing a wide range of non-stationary signals and can quantify statistical local self-similarity throughout the time-sequence ultrasound images. The locally characterized boundaries of the fBm segmented LV were further iteratively refined using global information by means of second-order moments. The method is benchmarked using synthetic 3D echocardiography time-sequence ultrasound images for normal and different ischemic cardiomyopathy, and results compared with state-of-the-art LV segmentation. Furthermore, preliminary results on real data from canine cases is presented.

Omar S. Al-Kadi, Allen Lu, Albert J. Sinusas, James S. Duncan

Context Aware 3D Fully Convolutional Networks for Coronary Artery Segmentation

Cardiovascular disease caused by coronary artery disease (CAD) is one of the most common causes of death worldwide. Coronary artery segmentation has attracted increasing attention since it is useful for better visualization and diagnosis. Conventional lumen segmentation methods basically describe vessels by a rough tubular model, thus presenting inferiority on abnormal vascular structures and failing to distinguish exact coronary arteries from vessel-like structures. In this paper, we propose a context aware 3D fully convolutional network (FCN) for vessel enhancement and segmentation in coronary computed tomography angiography (CTA) volumes. Combining the superior capacity of CNN in extracting discriminative features and satisfactory suppression of vessel-like structures by spatial prior knowledge embedded, the proposed approach significantly outperforms conventional Hessian vesselness based approach on a dataset of 50 coronary CTA volumes.

Yongjie Duan, Jianjiang Feng, Jiwen Lu, Jie Zhou

Learning Associations Between Clinical Information and Motion-Based Descriptors Using a Large Scale MR-derived Cardiac Motion Atlas

The availability of large scale databases containing imaging and non-imaging data, such as the UK Biobank, represents an opportunity to improve our understanding of healthy and diseased bodily function. Cardiac motion atlases provide a space of reference in which the motion fields of a cohort of subjects can be directly compared. In this work, a cardiac motion atlas is built from cine MR data from the UK Biobank ( $$\approx $$ 6000 subjects). Two automated quality control strategies are proposed to reject subjects with insufficient image quality. Based on the atlas, three dimensionality reduction algorithms are evaluated to learn data-driven cardiac motion descriptors, and statistical methods used to study the association between these descriptors and non-imaging data. Results show a positive correlation between the atlas motion descriptors and body fat percentage, basal metabolic rate, hypertension, smoking status and alcohol intake frequency. The proposed method outperforms the ability to identify changes in cardiac function due to these known cardiovascular risk factors compared to ejection fraction, the most commonly used descriptor of cardiac function. In conclusion, this work represents a framework for further investigation of the factors influencing cardiac health.

Esther Puyol-Antón, Bram Ruijsink, Hélène Langet, Mathieu De Craene, Paolo Piro, Julia A. Schnabel, Andrew P. King

Computational Modelling of Electro-Mechanical Coupling in the Atria and Its Changes During Atrial Fibrillation

Atrial fibrillation (AF) is a very common, multifaceted disease that affects atrial structure as well as electrophysiological and biomechanical function. However, the mechanistic links between structural and functional factors underlying AF are poorly understood. To explore these mechanisms, a 3D atrial electro-mechanical (EM) model was developed that includes 3D atrial geometry based on the Visible Female dataset, rule-based fibre orientation, CRN human atrial electrophysiology model and activation-based mechanical contraction model. Electrical activation in the 3D atria was simulated under control condition and two AF scenarios: sinus rhythm (SR), functional re-entry in the right atrium (RA) and structural re-entry around fibrotic patches in the left atrium (LA). Fibrosis distributions were obtained from patient LGE MRI data. In both AF scenarios, re-entrant behaviours led to substantial reductions in the displacement at peak contraction compared to SR. Specifically, high-frequency re-entry led to a decrease in maximal displacement from 6.8 to 6.1 mm in the posterior RA, and a larger decrease from 7.8 to 4.5 mm in the LA in the presence of fibrotic patches. The simulated displacement values agreed with available clinical data. In conclusion, the novel model of EM coupling in the 3D human atria provided new insights into the mechanistic links between atrial electrics and mechanics during normal activation and re-entry sustaining AF. Re-entry in the RA and LA resulted in weaker contractions compared to SR, with additional effect of fibrosis on the atrial wall stiffness further reducing the contraction.

Sofia Monaci, David Nordsletten, Oleg Aslanidi

High Throughput Computation of Reference Ranges of Biventricular Cardiac Function on the UK Biobank Population Cohort

The exploitation of large-scale population data has the potential to improve healthcare by discovering and understanding patterns and trends within this data. To enable high throughput analysis of cardiac imaging data automatically, a pipeline should comprise quality monitoring of the input images, segmentation of the cardiac structures, assessment of the segmentation quality, and parsing of cardiac functional indexes. We present a fully automatic, high throughput image parsing workflow for the analysis of cardiac MR images, and test its performance on the UK Biobank (UKB) cardiac dataset. The proposed pipeline is capable of performing end-to-end image processing including: data organisation, image quality assessment, shape model initialisation, segmentation, segmentation quality assessment, and functional parameter computation; all without any user interaction. To the best of our knowledge, this is the first paper tackling the fully automatic 3D analysis of the UKB population study, providing reference ranges for all key cardiovascular functional indexes, from both left and right ventricles of the heart. We tested our workflow on a reference cohort of 800 healthy subjects for which manual delineations, and reference functional indexes exist. Our results show statistically significant agreement between the manually obtained reference indexes, and those automatically computed using our framework.

Rahman Attar, Marco Pereañez, Ali Gooya, Xènia Albà, Le Zhang, Stefan K. Piechnik, Stefan Neubauer, Steffen E. Petersen, Alejandro F. Frangi

Lumen Segmentation of Aortic Dissection with Cascaded Convolutional Network

For the diagnosis and treatment of aortic dissection, where blood flows in between the layers of the aortic wall, the segmentation of true and false lumens is necessary. This is a challenging task because the intimal flap separating true and false lumens is thin, discontinuous and has a complex shape. In this paper we formulate lumen segmentation of aortic dissection as the extraction of aortic adventitia (the contour of aorta) and intima (the contour of true lumen). To this end, we propose a cascaded convolutional network for contour extraction on 2-D cross-section images, and then construct a 3-D adventitia and intima shape model. We performed a five-fold cross-validation on 45 aortic dissection CT volumes. The proposed method demonstrated a good performance for both aorta and true lumen segmentation, and the mean Dice similarity coefficient was 0.989 for aorta and 0.925 for true lumen.

Ziyan Li, Jianjiang Feng, Zishun Feng, Yunqiang An, Yang Gao, Bin Lu, Jie Zhou

A Vessel-Focused 3D Convolutional Network for Automatic Segmentation and Classification of Coronary Artery Plaques in Cardiac CTA

The segmentation and classification of atherosclerotic plaque (AP) are of great importance in the diagnosis and treatment of coronary artery disease. Although the constitution of AP can be assessed through a contrast-enhanced coronary computed tomography angiography (CCTA), the interpretation of CCTA scans is time-consuming and tedious for radiologists. Automation of AP segmentation is highly desired for clinical applications and further researches. However, it is difficult due to the extreme unbalance of voxels, similar appearance between some plaques and background tissues, and artefacts. In this paper, we propose a vessel-focused 3D convolutional network for automatic segmentation of AP including three subtypes: calcified plaques (CAP), non-calcified plaques (NCAP) and mixed calcified plaques (MCAP). We first extract the coronary arteries from the CT volumes; then we reform the artery segments into straightened volumes; finally, a 3D vessel-focused convolutional neural network is employed for plaque segmentation. The proposed method is trained and tested on a dataset of multi-phase CCTA volumes of 25 patients. We further investigate the effect of artery straightening through a comparison experiment, in which the network is trained on original CT volumes. Results show that by artery extraction and straightening, the training time is reduced by 40% and the segmentation performance of non-calcified plaques and mixed calcified plaques gains significantly. The proposed method achieves dice scores of 0.83, 0.73 and 0.68 for CAP, NCAP and MCAP respectively on the test set, which shows potential value for clinical application.

Jiang Liu, Cheng Jin, Jianjiang Feng, Yubo Du, Jiwen Lu, Jie Zhou

Semi-automated Image Segmentation of the Midsystolic Left Ventricular Mitral Valve Complex in Ischemic Mitral Regurgitation

Ischemic mitral regurgitation (IMR) is primarily a left ventricular disease in which the mitral valve is dysfunctional due to ventricular remodeling after myocardial infarction. Current automated methods have focused on analyzing the mitral valve and left ventricle independently. While these methods have allowed for valuable insights into mechanisms of IMR, they do not fully integrate pathological features of the left ventricle and mitral valve. Thus, there is an unmet need to develop an automated segmentation algorithm for the left ventricular mitral valve complex, in order to allow for a more comprehensive study of this disease. The objective of this study is to generate and evaluate segmentations of the left ventricular mitral valve complex in pre-operative 3D transesophageal echocardiography using multi-atlas label fusion. These patient-specific segmentations could enable future statistical shape analysis for clinical outcome prediction and surgical risk stratification. In this study, we demonstrate a preliminary segmentation pipeline that achieves an average Dice coefficient of 0.78 ± 0.06.

Ahmed H. Aly, Abdullah H. Aly, Mahmoud Elrakhawy, Kirlos Haroun, Luis Prieto-Riascos, Robert C. Gorman, Natalie Yushkevich, Yoshiaki Saito, Joseph H. Gorman, Robert C. Gorman, Paul A. Yushkevich, Alison M. Pouch

Atrial Scar Segmentation via Potential Learning in the Graph-Cut Framework

Late Gadolinium Enhancement Magnetic Resonance Imaging (LGE MRI) emerges as a routine scan for patients with atrial fibrillation (AF). However, due to the low image quality automating the quantification and analysis of the atrial scars is challenging. In this study, we proposed a fully automated method based on the graph-cut framework, where the potential of the graph is learned on a surface mesh of the left atrium (LA), using an equidistant projection and a deep neural network (DNN). For validation, we employed 100 datasets with manual delineation. The results showed that the performance of the proposed method was improved and converged with respect to the increased size of training patches, which provide important features of the structural and texture information learned by the DNN. The segmentation could be further improved when the contribution from the t-link and n-link is balanced, thanks to the inter-relationship learned by the DNN for the graph-cut algorithm. Compared with the existing methods which mostly acquired an initialization from manual delineation of the LA or LA wall, our method is fully automated and has demonstrated great potentials in tackling this task. The accuracy of quantifying the LA scars using the proposed method was 0.822, and the Dice score was 0.566. The results are promising and the method can be useful in diagnosis and prognosis of AF.

Lei Li, Guang Yang, Fuping Wu, Tom Wong, Raad Mohiaddin, David Firmin, Jenny Keegan, Lingchao Xu, Xiahai Zhuang

4D Cardiac Motion Modeling Using Pair-Wise Mesh Registration

In this paper, we present a novel method for the real-time cardiac motion compensation. Our method generates interpolated cardiac motion using segmented mesh models from preoperative 3D+T computed tomography angiography (CTA). We propose a pair-wise mesh registration technique for building correspondence and interpolating the control points over a cardiac cycle. The key contribution of this work is a rapid creation of a deformation field through a concise mathematical formulation while maintaining desired properties. These are $$C^2$$ continuity, invertibility, incompressibility of cardiac structure and capability to handling large deformation. And we evaluated the proposed method using different conditions, such as deformation resolution, temporal sampling rates, and template model selection.

Siyeop Yoon, Stephen Baek, Deukhee Lee

ISACHI: Integrated Segmentation and Alignment Correction for Heart Images

We address the problem of cardiovascular shape representation from misaligned Cardiovascular Magnetic Resonance (CMR) images. An accurate 3D representation of the heart geometry allows for robust metrics to be calculated for multiple applications, from shape analysis in populations to precise description and quantification of individual anatomies including pathology. Clinical CMR relies on the acquisition of heart images at different breath holds potentially resulting in a misaligned stack of slices. Traditional methods for 3D reconstruction of the heart geometry typically rely on alignment, segmentation and reconstruction independently. We propose a novel method that integrates simultaneous alignment and segmentation refinements to realign slices producing a spatially consistent arrangement of the slices together with their segmentations fitted to the image data.

Benjamin Villard, Ernesto Zacur, Vicente Grau

3D LV Probabilistic Segmentation in Cardiac MRI Using Generative Adversarial Network

Cardiac magnetic resonance imaging (MRI) is one of the most useful techniques to understand and measure cardiac functions. Given the MR image data, segmentation of the left ventricle (LV) myocardium is the most common task to be addressed for recovering and studying LV wall motion. However, most of segmentation methods heavily rely on the imaging appearance for extracting the myocardial contours (epi- and endo-cardium). These methods cannot guarantee a consistent volume of the heart wall during cardiac cycle in reconstructed 3D LV wall models, which is contradictory to the assumption of approximately constant myocardial tissue. In the paper, we propose a probability-based segmentation method to estimate the probabilities of boundary pixels in the trabeculated region near the solid wall belonging to blood or muscle. It helps avoid artifactually moving the endocardium boundary inward during systole, as commonly happens with simple threshold-based segmentation methods. Our method takes s stack of 2D cine MRI slices as input, and produces the 3D probabilistic segmentation of the heart wall using a generative adversarial network (GAN). Based on numerical experiments, our proposed method outperformed the baseline method in terms of evaluation metrics on a synthetic dataset. Moreover, we achieved very good quality reconstructed results with on a real 2D cine MRI dataset (there is no truly independent 3D ground truth). The proposed approach helps to achieve better understanding cardiovascular motion. Moreover, it is the first attempt to use probabilistic segmentation of LV myocardium for 3D heart wall reconstruction from 2D cardiac cine MRI data, to the best of our knowledge.

Dong Yang, Bo Liu, Leon Axel, Dimitris Metaxas

A Two-Stage U-Net Model for 3D Multi-class Segmentation on Full-Resolution Cardiac Data

Deep convolutional neural networks (CNNs) have achieved state-of-the-art performances for multi-class segmentation of medical images. However, a common problem when dealing with large, high resolution 3D data is that the volumes input into the deep CNNs has to be either cropped or downsampled due to limited memory capacity of computing devices. These operations can lead to loss of resolution and class imbalance in the input data batches, thus downgrade the performances of segmentation algorithms. Inspired by the architecture of image super-resolution CNN (SRCNN), we propose a two-stage modified U-Net framework that simultaneously learns to detect a ROI within the full volume and to classify voxels without losing the original resolution. Experiments on a variety of multi-modal 3D cardiac images have demonstrated that this framework shows better segmentation performances than state-of-the-art Deep CNNs with trained with the same similarity metrics.

Chengjia Wang, Tom MacGillivray, Gillian Macnaught, Guang Yang, David Newby

Centreline-Based Shape Descriptors of the Left Atrial Appendage in Relation with Thrombus Formation

The majority of thrombi in non-valvular atrial fibrillation (AF) patients are formed in the left atrial appendage (LAA). The shape of the LAA is highly variable and complex morphologies seem to favour the development of thrombus since they may induce blood stasis. Nevertheless, the relation between LAA shape and risk of thrombus has not been rigorously studied due to the lack of appropriate imaging data and robust tools to characterize LAA morphology. The main goal of this work was to automatically extract simple LAA morphological descriptors and study their relation with the presence of thrombus. LAA shape characterization was based on the computation of its centreline combining a heat transfer-derived distance map and a marching algorithm, once it LAA was segmented from 3D medical images. From the LAA centreline, several morphological descriptors were derived such as its length, tortuosity and major bending angles, among others. Other LAA parameters such as surface area, volume and ostium shape parameters were also obtained. A total of 71 LAA geometries from AF patients were analysed; 33 of them with a history of a thromboembolic event. We performed a statistical analysis to identify morphological descriptors showing differences between patients with and without thrombus history. Ostium size and centreline length were significatively different in both cohorts, with larger average values in thromboembolic cases, which could be related to slower blood flow velocities within the LAA. Additionally, some of the obtained centreline-based LAA parameters could be used for a better planning of LAA occluder implantations.

Ibai Genua, Andy L. Olivares, Etelvino Silva, Jordi Mill, Alvaro Fernandez, Ainhoa Aguado, Marta Nuñez-Garcia, Tom de Potter, Xavier Freixa, Oscar Camara

3D Atrial Segmentation Challenge


Automatic 3D Atrial Segmentation from GE-MRIs Using Volumetric Fully Convolutional Networks

In this paper, we propose an approach for automatic 3D atrial segmentation from Gadolinium-enhanced MRIs based on volumetric fully convolutional networks. The entire framework consists of two networks, the first network is to roughly locate the atrial center based on a low-resolution down-sampled version of the input and cut out a fixed size area that covers the atrial cavity, leaving out other pixels irrelevant to reduce memory consumption, and the second network is to precisely segment atrial cavity from the cropped sub-regions obtained from last step. Both two networks are trained end-to-end from scratch using 2018 Atrial Segmentation Challenge ( ) dataset which contains 100 GE-MRIs for training, and our method achieves satisfactory segmentation accuracy, up to 0.932 in Dice Similarity Coefficient score evaluated on the 54 testing samples, which ranks 1st among all participants.

Qing Xia, Yuxin Yao, Zhiqiang Hu, Aimin Hao

Automatically Segmenting the Left Atrium from Cardiac Images Using Successive 3D U-Nets and a Contour Loss

Radiological imaging offers effective measurement of anatomy, which is useful in disease diagnosis and assessment. Previous study [1] has shown that the left atrial wall remodeling can provide information to predict treatment outcome in atrial fibrillation. Nevertheless, the segmentation of the left atrial structures from medical images is still very time-consuming. Current advances in neural network may help creating automatic segmentation models that reduce the workload for clinicians. In this preliminary study, we propose automated, two-stage, three-dimensional U-Nets with convolutional neural network, for the challenging task of left atrial segmentation. Unlike previous two-dimensional image segmentation methods, we use 3D U-Nets to obtain the heart cavity directly in 3D. The dual 3D U-Net structure consists of, a first U-Net to coarsely segment and locate the left atrium, and a second U-Net to accurately segment the left atrium under higher resolution. In addition, we introduce a Contour loss based on additional distance information to adjust the final segmentation. We randomly split the data into training datasets (80 subjects) and validation datasets (20 subjects) to train multiple models, with different augmentation setting. Experiments show that the average Dice coefficients for validation datasets are around 0.91–0.92, the sensitivity around 0.90–0.94 and the specificity 0.99. Compared with traditional Dice loss, models trained with Contour loss in general offer smaller Hausdorff distance with similar Dice coefficient, and have less connected components in predictions. Finally, we integrate several trained models in an ensemble prediction to segment testing datasets.

Shuman Jia, Antoine Despinasse, Zihao Wang, Hervé Delingette, Xavier Pennec, Pierre Jaïs, Hubert Cochet, Maxime Sermesant

Fully Automated Left Atrium Cavity Segmentation from 3D GE-MRI by Multi-atlas Selection and Registration

This paper presents a fully automated method to segment the complex left atrial (LA) cavity, from 3D Gadolinium-enhanced magnetic resonance imaging (GE-MRI) scans. The proposed method consists of four steps: (1) preprocessing to convert the original GE-MRI to a probability map, (2) atlas selection to match the atlases to the target image, (3) multi-atlas registration and fusion, and (4) level-set refinement. The method was evaluated on the datasets provided by the MICCAI 2018 STACOM Challenge with 100 dataset for training. Compared to manual annotation, the proposed method achieved an average Dice overlap index of 0.88.

Mengyun Qiao, Yuanyuan Wang, Rob J. van der Geest, Qian Tao

Pyramid Network with Online Hard Example Mining for Accurate Left Atrium Segmentation

Accurately segmenting left atrium in MR volume can benefit the ablation procedure of atrial fibrillation. Traditional automated solutions often fail in relieving experts from the labor-intensive manual labeling. In this paper, we propose a deep neural network based solution for automated left atrium segmentation in gadolinium-enhanced MR volumes with promising performance. We firstly argue that, for this volumetric segmentation task, networks in 2D fashion can present great superiorities in time efficiency and segmentation accuracy than networks with 3D fashion. Considering the highly varying shape of atrium and the branchy structure of associated pulmonary veins, we propose to adopt a pyramid module to collect semantic cues in feature maps from multiple scales for fine-grained segmentation. Also, to promote our network in classifying the hard examples, we propose an Online Hard Negative Example Mining strategy to identify voxels in slices with low classification certainties and penalize the wrong predictions on them. Finally, we devise a competitive training scheme to further boost the generalization ability of networks. Extensively verified on 20 testing volumes, our proposed framework achieves an average Dice of $$92.83\%$$ in segmenting the left atria and pulmonary veins.

Cheng Bian, Xin Yang, Jianqiang Ma, Shen Zheng, Yu-An Liu, Reza Nezafat, Pheng-Ann Heng, Yefeng Zheng

Combating Uncertainty with Novel Losses for Automatic Left Atrium Segmentation

Segmenting left atrium in MR volume holds great potentials in promoting the treatment of atrial fibrillation. However, the varying anatomies, artifacts and low contrasts among tissues hinder the advance of both manual and automated solutions. In this paper, we propose a fully-automated framework to segment left atrium in gadolinium-enhanced MR volumes. The region of left atrium is firstly automatically localized by a detection module. Our framework then originates with a customized 3D deep neural network to fully explore the spatial dependency in the region for segmentation. To alleviate the risk of low training efficiency and potential overfitting, we enhance our deep network with the transfer learning and deep supervision strategy. Main contribution of our network design lies in the composite loss function to combat the boundary ambiguity and hard examples. We firstly adopt the Overlap loss to encourage network reduce the overlap between the foreground and background and thus sharpen the predictions on boundary. We then propose a novel Focal Positive loss to guide the learning of voxel-specific threshold and emphasize the foreground to improve classification sensitivity. Further improvement is obtained with an recursive training scheme. With ablation studies, all the introduced modules prove to be effective. The proposed framework achieves an average Dice of $$92.24\%$$ in segmenting left atrium with pulmonary veins on 20 testing volumes.

Xin Yang, Na Wang, Yi Wang, Xu Wang, Reza Nezafat, Dong Ni, Pheng-Ann Heng

Attention Based Hierarchical Aggregation Network for 3D Left Atrial Segmentation

Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. The atrial segmentation is essential for the understanding of the human atria structure which is vital to the AF treatment. In this paper, we propose a novel three-dimensional (3D) segmentation network combining hierarchical aggregation and attention mechanism for 3D left atrial segmentation, named attention based hierarchical aggregation network (HAANet). In our network, the shallow and deep feature fusion capability of encoder-decoder convolutional neural networks is enhanced through hierarchical aggregation. Besides, attention mechanism is adopted to promote the ability of extracting efficient features. Experimental results demonstrate the HAANet can produce good results for 3D left atrial segmentation and the dice score of our HAANet reaches 92.30.

Caizi Li, Qianqian Tong, Xiangyun Liao, Weixin Si, Yinzi Sun, Qiong Wang, Pheng-Ann Heng

Segmentation of the Left Atrium from 3D Gadolinium-Enhanced MR Images with Convolutional Neural Networks

Gadolinium contrast agents are used in a third of all magnetic resonance scans to study the extent of fibrosis across the left atria in patients with atrial fibrillation. Direct segmentation of the atrial heart chambers from these 3D gadolinium-enhanced magnetic resonance images (GE-MRI) is very demanding due to the low contrast between the atrial tissue and background. Current automatic segmentation methods delineate the left atrium on auxiliary bright-blood MRI scans, which need to be registered to GE-MRI in an additional, potentially error-prone step. Yet, it could render extremely useful to obtain direct segmentation on GE-MRI to simultaneously estimate the left atrium anatomy and the extent of fibrosis. In this work, we present a deep learning approach which is able to segment the left atrium from 3D GE-MRI. 100 data sets provided by the MICCAI 2018 Atrial Segmentation Challenge have been used to train and test deep convolutional neural networks (CNN), which follow a 2D architecture with deep supervision (train-validation-test 70-5-25). After performing a four fold cross validation, the network achieved a mean dice of 0.8945. Within the scope of the test phase of the challenge, we trained the network on 100 data sets, predicted novel segmentations on the official test set and, according to the leaderboard, achieved a final score of 0.888.

Chandrakanth Jayachandran Preetha, Shyamalakshmi Haridasan, Vahid Abdi, Sandy Engelhardt

V-FCNN: Volumetric Fully Convolution Neural Network for Automatic Atrial Segmentation

Atrial Fibrillation (AF) is a common electro-physiological cardiac disorder that causes changes in the anatomy of the atria. A better characterization of these changes is desirable for the definition of clinical biomarkers. There is thus a need for its fully automatic segmentation from clinical images. This work presents an architecture based on 3D-convolution kernels, a Volumetric Fully Convolution Neural Network (V-FCNN), able to segment the entire atrial anatomy in a one-shot from high-resolution images ( $$640\times 640$$ pixels). A loss function based on the mixture of both Mean Square Error (MSE) and Dice Loss (DL) is used, in an attempt to combine the ability to capture the bulk shape as well as the reduction of local errors caused by over-segmentation.Results demonstrate a good performance in the middle region of the atria along with the challenges impact of capturing the pulmonary veins variability or valve plane identification that separates the atria to the ventricle. Despite the need to reduce the original image resolution to fit into Graphics Processing Unit (GPU) hardware constraints, $$92.5\%$$ and $$85.1\%$$ were obtained respectively in the 2D and 3D Dice metric in 54 test patients (4752 atria test slices in total), making the V-FCNN a reasonable model to be used in clinical practice.

Nicoló Savioli, Giovanni Montana, Pablo Lamata

Ensemble of Convolutional Neural Networks for Heart Segmentation

Training an ensemble of convolutional neural networks requires much computational resources for a large set of high-resolution medical 3D scans because deep representation requires many parameters and layers. In this study, 100 3D late gadolinium-enhanced (LGE)-MRIs with a spatial resolution of 0.625 mm × 0.625 mm × 0.625 mm from patients with atrial fibrillation were utilized. To contain cost of the training, down-sampling of images, transfer learning and ensemble of network’s past weights were deployed. This paper proposes an image processing stage using down-sampling and contrast limited adaptive histogram equalization, a network training stage using a cyclical learning rate schedule, and a testing stage using an ensemble. While this method achieves reasonable segmentation accuracy with the median of the Dice coefficients at 0.87, this method can be used on a computer with a GPU that has a Kepler architecture and at least 3 GB memory.

Wilson Fok, Kevin Jamart, Jichao Zhao, Justin Fernandez

Multi-task Learning for Left Atrial Segmentation on GE-MRI

Segmentation of the left atrium (LA) is crucial for assessing its anatomy in both pre-operative atrial fibrillation (AF) ablation planning and post-operative follow-up studies. In this paper, we present a fully automated framework for left atrial segmentation in gadolinium-enhanced magnetic resonance images (GE-MRI) based on deep learning. We propose a fully convolutional neural network and explore the benefits of multi-task learning for performing both atrial segmentation and pre/post ablation classification. Our results show that, by sharing features between related tasks, the network can gain additional anatomical information and achieve more accurate atrial segmentation, leading to a mean Dice score of 0.901 on a test set of 20 3D MRI images. Code of our proposed algorithm is available at .

Chen Chen, Wenjia Bai, Daniel Rueckert

Left Atrial Segmentation Combining Multi-atlas Whole Heart Labeling and Shape-Based Atlas Selection

Segmentation of the left atria (LA) from late gadolinium enhanced magnetic resonance imaging (LGE-MRI) is challenging since atrial borders are not easily distinguishable in the images. We propose a method based on multi-atlas whole heart segmentation and shape modeling of the LA. In the training phase we first construct whole heart LGE-MRI atlases and build a principal component analysis (PCA) model able to capture the high variability of the LA shapes. All atlases are clustered according to their LA shape using an unsupervised clustering method which additionally outputs the most representative case in each cluster. All cluster representatives are registered to the target image and ranked using conditional entropy. A small subset of the most similar representatives is used to find LA shapes with similar morphology in the training set that are used to obtain the final LA segmentation. We tested our approach using 80 LGE-MRI data for training and 20 LGE-MRI data for testing obtaining a Dice score of $$0.842 \pm 0.049$$ .

Marta Nuñez-Garcia, Xiahai Zhuang, Gerard Sanroma, Lei Li, Lingchao Xu, Constantine Butakoff, Oscar Camara

Deep Learning Based Method for Left Atrial Segmentation in GE-MRI

Understanding the anatomical structure of left atrial (LA) is crucial for clinical treatment of atrial fibrillation (AF). Gadolinium Enhanced Magnetic Resonance Imaging (GE-MRI) provides clarity images of LA structure. However, the most of LA structure analysis on GE-MRI studies are based on subjective manual segmentation. An efficient and objective segmentation method in GE-MRI is highly demanded. Although deep learning based method has achieved great success on some medical image segmentations, solving LA segmentation through deep learning is still an unsatisfied field. In this paper, we handle this unmet clinical need by exploring two convolutional neural networks (CNNs) structures, fully convolutional network (FCN) and U-Net, to improve the accuracy and efficiency of LA segmentation. Both models were trained and evaluated on GE-MRI dataset provided by 2018 atrial segmentation challenge. The results show that FCN-based LA automatic segmentation method achieves Dice score over 82%; U-Net method achieves Dice score over 83%.

Yashu Liu, Yangyang Dai, Cong Yan, Kuanquan Wang

Dilated Convolutions in Neural Networks for Left Atrial Segmentation in 3D Gadolinium Enhanced-MRI

Segmentation of the left atrial chamber and assessing its morphology, are essential for improving our understanding of atrial fibrillation, the most common type of cardiac arrhythmia. Automation of this process in 3D gadolinium enhanced-MRI (GE-MRI) data is desirable, as manual delineation is time-consuming, challenging and observer-dependent. Recently, deep convolutional neural networks (CNNs) have gained tremendous traction and achieved state-of-the-art results in medical image segmentation. However, it is difficult to incorporate local and global information without using contracting (pooling) layers, which in turn reduces segmentation accuracy for smaller structures. In this paper, we propose a 3D CNN for volumetric segmentation of the left atrial chamber in LGE-MRI. Our network is based on the well known U-Net architecture. We employ a 3D fully convolutional network, with dilated convolutions in the lowest level of the network, and residual connections between encoder blocks to incorporate local and global knowledge. The results show that including global context through the use of dilated convolutions, helps in domain adaptation, and the overall segmentation accuracy is improved in comparison to a 3D U-Net.

Sulaiman Vesal, Nishant Ravikumar, Andreas Maier

A Semantic-Wise Convolutional Neural Network Approach for 3-D Left Atrium Segmentation from Late Gadolinium Enhanced Magnetic Resonance Imaging

Several studies suggest that the assessment of viable left atrial (LA) tissue is a relevant information to support catheter ablation in atrial fibrillation (AF). Late gadolinium enhanced magnetic resonance imaging (LGE MRI) is a new emerging technique which is employed for the non-invasive quantification of LA fibrotic tissue. The analysis of LGE MRI relies on manual tracing of LA boundaries. This procedure is time-consuming and prone to high inter-observer variability given the different degrees of observers’ experience, LA wall thickness and data resolution. Therefore, an automatic approach for the LA wall detection would be highly desirable. This work focuses on the design and development of a semantic-wise convolutional neural network based on the successful architecture U-Net (U-SWCNN). Batch normalization, early stopping and parameter initializers consistent with the activation functions chosen were used; a loss function based on the Dice coefficient was employed. The U-SWCNN was trained end-to-end with the 3-D data available from the 2018 Atrial Segmentation Challenge. The training was completed using 80 LGE MRI data and a post-processing step based on the 3-D morphology was then applied. After the post-processing step, the average Dice coefficient on the validation set (20 LGE MRI data) was 0.911, while on the test set (54 LGE MRI data) was 0.898.

Davide Borra, Alessandro Masci, Lorena Esposito, Alice Andalò, Claudio Fabbri, Cristiana Corsi

Left Atrial Segmentation in a Few Seconds Using Fully Convolutional Network and Transfer Learning

In this paper, we propose a fast automatic method that segments left atrial cavity from 3D GE-MRIs without any manual assistance, using a fully convolutional network (FCN) and transfer learning. This FCN is the base network of VGG-16, pre-trained on ImageNet for natural image classification, and fine tuned with the training dataset of the MICCAI 2018 Atrial Segmentation Challenge. It relies on the “pseudo-3D” method published at ICIP 2017, which allows for segmenting objects from 2D color images which contain 3D information of MRI volumes. For each $$n^{\text {th}}$$ slice of the volume to segment, we consider three images, corresponding to the $$(n-1)^{\text {th}}$$ , $$n^{\text {th}}$$ , and $$(n+1)^{\text {th}}$$ slices of the original volume. These three gray-level 2D images are assembled to form a 2D RGB color image (one image per channel). This image is the input of the FCN to obtain a 2D segmentation of the $$n^{\text {th}}$$ slice. We process all slices, then stack the results to form the 3D output segmentation. With such a technique, the segmentation of the left atrial cavity on a 3D volume takes only a few seconds. We obtain a Dice score of 0.92 both on the training set in our experiments before the challenge, and on the test set of the challenge.

Élodie Puybareau, Zhou Zhao, Younes Khoudli, Edwin Carlinet, Yongchao Xu, Jérôme Lacotte, Thierry Géraud

Convolutional Neural Networks for Segmentation of the Left Atrium from Gadolinium-Enhancement MRI Images

This paper introduces a left atrial segmentation pipeline that utilises a deep neural network for learning segmentations of the LA from Gadolinium enhancement magnetic resonance images (GE-MRI). The trainable fully-convolutional neural network consists of an encoder network and a corresponding decoder network followed by a pixel-wise classification layer. The entire network has 17 convolutional layers, with the encoder network containing 5 convolutional layers, and the decoder network containing 11 convolution layers with 1 additional convolution layer in between. The training image database consisted of manually annotated GE-MRI images ( $$n=75$$ ). Dice scores of $$0.87 \pm 0.07$$ and $$0.80 \pm 0.12$$ were achieved on our test set ( $$n=25$$ ) and a multi-centre independent set using transfer learning, respectively. On the test set that was provided by the challenge ( $$n=54$$ ), a Dice score of 0.897 was achieved. We experimentally demonstrated the robustness of the proposed method as a segmentation pipeline for potential use in clinical research.

Coen de Vente, Mitko Veta, Orod Razeghi, Steven Niederer, Josien Pluim, Kawal Rhode, Rashed Karim

Mixture Modeling of Global Shape Priors and Autoencoding Local Intensity Priors for Left Atrium Segmentation

Difficult image segmentation problems, e.g., left atrium in MRI, can be addressed by incorporating shape priors to find solutions that are consistent with known objects. Nonetheless, a single multivariate Gaussian is not an adequate model in cases with significant nonlinear shape variation or where the prior distribution is multimodal. Nonparametric density estimation is more general, but has a ravenous appetite for training samples and poses serious challenges in optimization, especially in high dimensional spaces. Here, we propose a maximum-a-posteriori formulation that relies on a generative image model by incorporating both local intensity and global shape priors. We use deep autoencoders to capture the complex intensity distribution while avoiding the careful selection of hand-crafted features. We formulate the shape prior as a mixture of Gaussians and learn the corresponding parameters in a high-dimensional shape space rather than pre-projecting onto a low-dimensional subspace. In segmentation, we treat the identity of the mixture component as a latent variable and marginalize it within a generalized expectation-maximization framework. We present a conditional maximization-based scheme that alternates between a closed-form solution for component-specific shape parameters that provides a global update-based optimization strategy, and an intensity-based energy minimization that translates the global notion of a nonlinear shape prior into a set of local penalties. We demonstrate our approach on the left atrial segmentation from gadolinium-enhanced MRI, which is useful in quantifying the atrial geometry in patients with atrial fibrillation.

Tim Sodergren, Riddhish Bhalodia, Ross Whitaker, Joshua Cates, Nassir Marrouche, Shireen Elhabian

Left Ventricle Full Quantification Challenge


Left-Ventricle Quantification Using Residual U-Net

Estimating dimensional measurements of the left ventricle provides diagnostic values which can be used to assess cardiac health and identify certain pathologies. In this paper we describe our methodology of calculating measurements from left ventricle segmentations automatically generated using deep learning. We use a U-net convolutional neural network architecture built from residual units to segment the left ventricle and then process these segmentations to estimate the area of the cavity and myocardium, the dimensions of the cavity, and the thickness of the myocardium. Determining if an image is part of the diastolic or systolic portion of the cardiac cycle is done by analysing the cavity volume. The quality of our results are dependent on our training regime where we have generated a large derivative dataset by augmenting the original images with free-form deformations. Our expanded training set, in conjunction with simple affine image transforms, creates a sufficiently large training population to prevent over-fitting of the network while still creating an accurate and robust segmentation network. Assessing our method on the STACOM18 LVQuan challenge dataset we find that it significantly outperforms the previously published state-of-the-art on a 5-fold validation all tasks considered.

Eric Kerfoot, James Clough, Ilkay Oksuz, Jack Lee, Andrew P. King, Julia A. Schnabel

Left Ventricle Full Quantification Using Deep Layer Aggregation Based Multitask Relationship Learning

Left ventricle full quantification is important in the assessment of cardiac functionality and diagnosis of cardiac diseases, but is also challenging due to the sample variability and label correlations. In this paper, we propose a deep-learning based approach for left ventricle full quantification, including 11 indices regression and cardiac phase recognition. We utilize Deep Layer Aggregation as backbone, perform 11 indices regression simultaneously supervised by multitask relationship loss, and then derive the cardiac phase by searching maximum and minimum frame from polynomial-fitted cavity area. Experiments demonstrate the superiority of the proposed method in performance.

Jiahui Li, Zhiqiang Hu

Convexity and Connectivity Principles Applied for Left Ventricle Segmentation and Quantification

We propose an unsupervised method for MRI image segmentation, and global and regional shape quantification, based on pixel labeling using image analysis, connectivity constraints and near convex region requirements for the LV cavity and the epicardium. The proposed method is developed in the framework of the MICCAI Left Ventricle Full Quantification Challenge. At first the LV cavity is approximately localized based on the strong intensity contrast in the myocardium region between the two ventricles (left and right). The requirement of a near convex connected component is then applied. The image intensity statistical parameters are extracted for three classes: LV cavity, myocardium and chest space. Even if the whole background is completely inhomogeneous, the application of topological, connectivity and shape constraints permits to extract in two steps the LV cavity and the myocardium. For the later two approaches are proposed: regularization using B-spline smoothing and adaptive region growing with boundary smoothing using Fourier coefficients. On the segmented images are measured the significant clinical global and regional shape LV indices. We consider that we have obtained good results on indices related to the endocardium for both Training and Test datasets. There is place for improvements concerning the myocardium global and regional shape indices.

Elias Grinias, Georgios Tziritas

Calculation of Anatomical and Functional Metrics Using Deep Learning in Cardiac MRI: Comparison Between Direct and Segmentation-Based Estimation

In this paper we propose a collection of left ventricle (LV) quantification methods using different versions of a common neural network architecture. In particular, we compare the accuracy obtained with direct calculation (regression) of the desired metrics, a segmentation network and a novel combined approach. We also introduce temporal dynamics through the use of a Long Short-Term Memory (LSTM) network. We train and evaluate our methods on MICCAI 2018 Left Ventricle Full Quantification Challenge dataset. We perform 5-fold cross-validation on the training dataset and compare our results with the state-of-the-art methods evaluated on the same dataset. In our experiments, segmentation-based methods outperform all the other variants as well as current state of the art. The introduction of LSTM does produces only minor improvements in accuracy. The novel segmentation/estimation network improves the results on estimation-only but does not reach the accuracy of segmentation-based metric calculation.

Hao Xu, Jurgen E. Schneider, Vicente Grau

Automated Full Quantification of Left Ventricle with Deep Neural Networks

Accurate cardiac left ventricle (LV) quantification is among the most clinically important and most frequently demanded tasks for identification and diagnosis of cardiac diseases and is of great interest in the research community of medical image analysis. However, it is still a task of great challenge due to the high variability of cardiac structure across subjects and the complexity of temporal dynamics of cardiac sequences. Full quantification of cardiac LV includes simultaneously quantifying, for every frame in the whole cardiac cycle, multiple types of cardiac indices, such as cavity and myocardium areas, regional wall thicknesses, LV dimension and cardiac phase. Accurate quantification of these indices will support comprehensive global and regional cardiac function assessment. In this paper, we propose a newly designed multitask learning network which combines the segmentation task and the cardiac indices quantification task. It comprises a segmentation network based on the U-net for image representation, and then followed by a simple convolutional neural network for feature extraction of cardiac indices, two parallel recurrent neural network models are then added for temporal dynamic modeling. Then we use multitask learning to capture the existing correlations among different tasks. Experiments of 5-fold validation results show that the proposed framework achieves high accurate prediction, with average mean absolute error of 173 mm $$^{2}$$ , 2.44 mm, 1.37 mm for areas, dimensions, RWT and phase error rate 7.8%.

Lihong Liu, Jin Ma, Jianzong Wang, Jing Xiao

ESU-P-Net: Cascading Network for Full Quantification of Left Ventricle from Cine MRI

Left ventricle (LV) quantification is of great clinical importance for diagnosing and monitoring cardiac diseases. Full quantification of LV indices includes: (1) two areas of LV cavity and myocardium, (2) six regional wall thicknesses (RWT), (3) three LV dimensions, and (4) phase identification (diastole or systole). However, due to the large variability in the object shape and imaging quality, it is time-consuming and user-dependent to quantify LV parameters manually. In this work, we propose a cascading deep neural network, including an enhanced supervision U-net followed a recurrent neural network (RNN) type of phase-prediction net called P-net, abbreviated as ESU-P-net, for full LV quantification in a fully automated manner. The proposed ESU-P-net framework is dedicated to the full quantification of LV for all four types of indices.Experiments on MR sequences of 145 subjects provided by MICCAI 2018 STACOM Challenge showed that the proposed network achieved highly accurate LV quantification, with an average mean absolute error (MAE) of 62 mm2, 1.14 mm, 0.96 mm for LV areas, RWT, dimensions, respectively, and an error rate of 8.0% for cardiac phase identification.

Wenjun Yan, Yuanyuan Wang, Shaoxiang Chen, Rob J. van der Geest, Qian Tao

Left Ventricle Full Quantification via Hierarchical Quantification Network

Automatic quantitative analysis of cardiac left ventricle (LV) function is one of challenging task for heart disease diagnosis. Four different parameters, i.e. regional wall thicknesses (RWT), area of myocardium and LV cavity, LV dimensions in different direction and cardiac phase, are used for evaluating the LV function. In this paper, we implemented a novel multi-task quantification network (HQNet) to simultaneously quantify the four different parameters. The network is mainly constituted by a customized convolutional neural network named Hierarchical convolutional neural network (HCNN) which includes different pyramid-like 3D convolution blocks with different kernel sizes for efficient feature embedding; and two long-short term memory (LSTM) networks for temporal modeling. Respecting inter-task correlations, our proposed network uses multi-task constraints for phase to improve the final estimation of phase. Selu activation function is selected instead of relu, which can bring better performance of model in experiments. Experiments on MR sequences of 145 patients show that HQNet achieves high accurate estimation by means of 7-fold cross validation. The mean absolute error (MAE) of average areas, RWT, dimensions are $$ 197\,{\text{mm}}^{2} ,1.51\,{\text{mm}},2.57\,{\text{mm}} $$ respectively. The error rate of phase classification is 9.8%. These results indicate that the approach we proposed has a promising performance to estimate all four parameters.

Guanyu Yang, Tiancong Hua, Chao Lu, Tan Pan, Xiao Yang, Liyu Hu, Jiasong Wu, Xiaomei Zhu, Huazhong Shu

Automatic Left Ventricle Quantification in Cardiac MRI via Hierarchical Refinement of High-Level Features by a Salient Perceptual Grouping Model

An accurate segmentation of the left ventricle (LV) from cardiac magnetic resonance imaging (MRI) provides reliable cardiac indexes such as the ventricular volume, the ejection fraction or regional wall thicknesses (RWT). This paper introduces an automated method to compute such indexes in 2D MRI slices from a semantic segmentation obtained in two steps. A first coarse segmentation is obtained by applying an encoder-decoder neural network architecture that assigns a probability value to each pixel. Afterwards, this segmentation is refined by a spatio-temporal saliency analysis. The method was evaluated in MR sequences of 175 subjects divided in two groups: training (145 subjects) and test (30 subjects). For the training data set, using a K-cross validation setup, the method achieves an average Pearson correlation coefficient of 0.98, 0.92, 0.95 and 0.75 with the set of indexes LV cavity, myocardium areas, cavity dimensions and region wall thicknesses, respectively, while classification of the cardiac phase yielded a rate of $$10.01\%$$ . For the same set of indexes, evaluated in the test dataset, an average Pearson correlation coefficient of 0.98, 0.87, 0.97 and 0.66 was obtained. Additionally, the cardiac phase classification error rate was $$9\%$$ . The method provides a reliable LV segmentation and quantification of cardiac indexes.

Angélica Atehortúa, Mireille Garreau, David Romo-Bucheli, Eduardo Romero

Cardiac MRI Left Ventricle Segmentation and Quantification: A Framework Combining U-Net and Continuous Max-Flow

Cardiac magnetic resonance imaging (MRI) is routinely used for cardiovascular disease diagnosis and therapy guidance. Left ventricle (LV) segmentation is typically required as a first step to quantify cardiac indices. In this work, we developed an automatic approach for LV segmentation and indices quantification of cardiac MRI. We employed a U-net convolutional neural network to generate LV segmentation probability maps. The initial probability maps were used to provide the labeling cost measurements of a continuous min-cut segmentation model and the final segmentation was regularized using image edge information. The continuous min-cut segmentation model was solved globally and exactly through convex relaxation and dual optimization on a GPU. We applied our approach to a clinical dataset of 45 subjects and achieved a mean DSC of $$89.4\pm 5.0\%$$ and average symmetric surface distance of $$0.81\pm 0.31$$ mm for LV myocardium segmentation. For LV indices quantification, we observed a mean absolute error of 114.8 mm $$^2$$ for LV cavity, 168.6 mm $$^2$$ for LV myocardium, $$\sim $$ 1.8 mm for LV cavity dimensions, and 1.2 $${\sim }$$ 1.6 mm for LV myocardium wall thickness measurements. These results suggest that our framework provide the potential for LV function quantification using cardiac MRI.

Fumin Guo, Matthew Ng, Graham Wright

Multi-estimator Full Left Ventricle Quantification Through Ensemble Learning

Cardiovascular disease accounts for 1 in every 4 deaths in United States. Accurate estimation of structural and functional cardiac parameters is crucial for both diagnosis and disease management. In this work, we develop an ensemble learning framework for more accurate and robust left ventricle (LV) quantification. The framework combines two 1st-level modules: direct estimation module and a segmentation module. The direct estimation module utilizes Convolutional Neural Network (CNN) to achieve end-to-end quantification. The CNN is trained by taking 2D cardiac images as input and cardiac parameters as output. The segmentation module utilizes a U-Net architecture for obtaining pixel-wise prediction of the epicardium and endocardium of LV from the background. The binary U-Net output is then analyzed by a separate CNN for estimating the cardiac parameters. We then employ linear regression between the 1st-level predictor and ground truth to learn a 2nd-level predictor that ensembles the results from 1st-level modules for the final estimation. Preliminary results by testing the proposed framework on the LVQuan18 dataset show superior performance of the ensemble learning model over the two base modules.

Jiasha Liu, Xiang Li, Hui Ren, Quanzheng Li

Left Ventricle Quantification Through Spatio-Temporal CNNs

Cardiovascular diseases are among the leading causes of death globally. Cardiac left ventricle (LV) quantification is known to be one of the most important tasks for the identification and diagnosis of such pathologies. In this paper, we propose a deep learning method that incorporates 3D spatio-temporal convolutions to perform direct left ventricle quantification from cardiac MR sequences. Instead of analysing slices independently, we process stacks of temporally adjacent slices by means of 3D convolutional kernels which fuse the spatio-temporal information, incorporating the temporal dynamics of the heart to the learned model. We show that incorporating such information by means of spatio-temporal convolutions into standard LV quantification architectures improves the accuracy of the predictions when compared with single-slice models, achieving competitive results for all cardiac indices and significantly breaking the state of the art [10] for cardiac phase estimation.

Alejandro Debus, Enzo Ferrante

Full Quantification of Left Ventricle Using Deep Multitask Network with Combination of 2D and 3D Convolution on 2D + t Cine MRI

Accurate quantification of left ventricle (LV) from cardiac image are valuable to evaluate ventricular function information such as stroke volume and ejection fraction. In this paper, we proposed a novel FCN architecture, which is trained in end-to-end manner, for full quantification of cardiac LV on 2D + t cine MR images. Considering 3D information as features for temporal modeling can improve performance of the model for temporal-related task. The proposed FCN with the alternate 3D-2D convolutional module addresses each sequence with assistance from adjacent sequences and shows the comparable results compared with the state-of-the-art method.

Yeonggul Jang, Sekeun Kim, Hackjoon Shim, Hyuk-Jae Chang


Additional information

Premium Partner

    Image Credits