Temporally Consistent Mitral Annulus Measurements from Sparse Annotations in Echocardiographic Videos

Gino E. Jansen, Mark J. Schuuring, Berto J. Bouma, Ivana Išgum · 2025

Read Temporally Consistent Mitral Annulus Measurements from Sparse Annotations in Echocardiographic Videos by Gino E. Jansen, Mark J. Schuuring, Berto J. Bouma, Ivana Išgum (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-12087.txt.

Introduction

Gino Jansen and co-authors track a heart valve ring in ultrasound videos in a 2025 paper. Doctors measure the mitral annulus over time, but marking every frame by hand is slow. Sparse marks leave gaps. Their method adds a self-supervised loss that keeps landmark spots steady between nearby frames.

Extra training tricks mimic a missing field of view, so the model learns to flag absent landmarks. Tests on public and private data sets show better tracking stability. Errors for a key motion measure and for ring size both stay low. The model also spots missing landmarks with very high accuracy, the authors report.

Catalog record

CREATORGino E. Jansen, Mark J. Schuuring, Berto J. Bouma, Ivana Išgum
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-12087

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Temporally Consistent Mitral Annulus Measurements from Sparse Annotations in Echocardiographic Videos Gino E. Jansen, Mark J. Schuuring, Berto J. Bouma, Ivana Išgum Abstract This work presents a novel approach to achieving temporally consistent mitral annulus landmark localization in echocardiography videos using sparse annotations. Our method introduces a self-supervised loss term that enforces temporal consistency between neighboring frames, which smooths the position of landmarks and enhances measurement accuracy over time. Additionally, we incorporate realistic field-of-view augmentations to improve the recognition of missing anatomical landmarks. We evaluate our approach on both a public and private dataset, and demonstrate significant improvements in Mitral Annular Plane Systolic Excursion (MAPSE) calculations and overall landmark tracking stability. The method achieves a mean absolute MAPSE error of 1.81 ± 0.14 mm, an annulus size error of 2.46 ± 0.31 mm, and a landmark localization error of 2.48 ± 0.07 mm. Finally, it achieves a 0.99 ROC-AUC for recognition of missing landmarks. a]Department of Biomedical Engineering & Physics, Amsterdam University Medical Center, Netherlands b]Informatics Institute, University of Amsterdam, Netherlands c]Amsterdam Cardiovascular Sciences, Netherlands d]Department of Biomedical Signals and Systems, University of Twente, Netherlands e]Department of Cardiology, Amsterdam University Medical Center, Netherlands f]Department of Radiology & Nuclear Medicine, Amsterdam University Medical Center, Netherlands keywords Echocardiography, Ultrasound, Landmark detection, Mitral valve annulus † † authorinfo: Further author information: Send correspondence to Gino E. Jansen. (E-mail: [email protected]) 1 INTRODUCTION The mitral valve plays a crucial role in cardiac function: It regulates blood flow between the left atrium and left ventricle. Accurate assessment of the mitral annulus – the fibrous ring supporting the mitral valve leaflets – is essential for diagnosing and managing various cardiac conditions. Changes in the diameter of the mitral annulus can indicate mitral regurgitation – unwanted backflow of blood from the left ventricle to the atrium. This condition often arises as a complication of coronary heart disease (CHD), a leading cause of death responsible for about one-third of all deaths in individuals over 35 years old [ 1 ] . Another key metric derived from the mitral annulus is the Mitral Annular Plane Systolic Excursion (MAPSE), which measures the longitudinal displacement of the mitral annulus during systole. MAPSE serves as a surrogate marker for left ventricular ejection fraction (LVEF), providing insights into left ventricular systolic function [ 2 ] , and it is of vital importance for diagnosing coronary heart disease. Reduced MAPSE values have been associated with increased mortality risk, independent of left ventricular ejection fraction [ 3 ] . Transthoracic echocardiography is the first-line imaging modality for evaluating the mitral annulus and calculating MAPSE. This requires localization of the two annulus landmarks at end-systole and end-diastole – time points that could be derived automatically if the anatomical landmarks are tracked accurately and consistently across sequential frames. Ensuring temporal consistency in these measurements is therefore crucial for providing meaningful clinical assessments. Recent advancements in deep learning have enabled supervised algorithms to perform landmark detection in medical images, learning from annotated examples [ 4 ] . However, manual frame-by-frame annotation of these landmarks is labor-intensive and prone to inconsistency across consecutive frames, highlighting the need for automated deep learning methods that ensure both accuracy and temporal consistency, without requiring a dense annotation strategy. Additionally, due to the limited field of view provided by ultrasound imaging, the target anatomy may often disappear because of unintended movements of the patient or operator during acquisition. An automatic landmark localization method needs to be able to account for this by recognizing when the target moves out of view. Several landmark detection methods have previously been developed for landmark detection in echocardiography, in particular for measuring left-ventricle wall thickness [ 5 , 6 , 7 ] . This task requires localization of landmarks along the boundaries of the septal and posterior wall in parasternal long-axis (PLAX) view. In these works, the authors employed fully convolutional networks – such as U-Net, DeeplabV3, or HRNet – to predict high-resolution heatmaps, whose centroids represent the predicted landmark location. These methods reach high performance on distance measures, but lack evaluations on temporal consistency, as their main focus lies with replicating the (sparse) annotations made by human observers. Several studies have addressed the challenge of temporally sparse annotations for deep-learning landmark detection in echocardiography videos [ 8 , 9 ] , but only for detecting landmarks on the inner wall of the left ventricle in PLAX-view. Furthermore, Smistad et al. [ 10 ] proposed a framework for landmark-tracking-based MAPSE calculation. This method firstly used a network that selects end-diastolic (ED) and end-systolic (ES) frames from a full video, then used another neural network to detect the annulus landmarks in ED, and propagated these landmarks over time with a third network that predicted the optical flow field between every two consecutive frames. Through landmark propagation, spatiotemporal relations were accounted for, but this also caused error accumulation. To incorporate spatiotemporal relations into automated landmark localization, we present a novel approach that incorporates consistency with neighboring frames, while requiring sparsely available landmark annotations only. Additionally, we propose training augmentations that improve the model’s ability to recognize missing (out-of-view) landmarks. We evaluate the method on two echocardiography datasets: a public dataset with left-heart segmentations of the apical 2- and 4-chamber view from which we extract mitral annulus landmarks; and a private dataset, which additionally includes apical 3-chamber views. 2 Data The datasets used in this study include the publicly available CAMUS dataset [ 11 ] , and a private dataset of echocardiography videos retrospectively collected from Amsterdam University Medical Center (AUMC). The annotations in both datasets include sparsely annotated left-ventricle contours, from which we extract the endpoints as target mitral annulus landmarks. The CAMUS dataset contains a 1,000 videos from 500 patients, with 49% of patients having a pathological ejection fraction below 45\% . Each patient study contains two videos showing the systolic phase of the A2C and A4C views, respectively, with each view containing a reference segmentation of the left ventricle for the end-diastolic and end-systolic frame. The AUMC dataset contains 293 videos from 104 patients with symptoms of angina and/or dyspnea, which include 100 apical 2-chamber (A2C), 92 apical 3-chamber (A3C), and 101 apical 4-chamber (A4C) views. For each video, several end-diastolic, end-systolic, and random intermediate frames were annotated, resulting in a total of 1,541 annotated frames. Figure 1 : Architecture of the proposed method. A ResNet-like fully convolutional neural network takes three consecutive frames as input, stacked in the channel dimension, and outputs classification and regression maps for mitral annulus landmarks. The blue convolutional blocks include GroupNorm, while other blocks omit normalization. ReLU is used as the activation function unless stated otherwise. Temporal consistency is enforced through an unsupervised loss that compares predicted displacements across neighboring frames, while supervised losses guide the regression and classification outputs. During test time, landmarks are computed as the weighted mean of regressed locations from each patch, using the respective probabilities from the classification map. 3 Method 3.1 Landmark detection Inspired by the landmark detection method proposed by Noothout et al. [ 12 ] , we employ a network that applies both classification and regression on a spatially down-sampled feature representation of the input image, as illustrated in Figure 1 . To enable landmark

Read the complete text