https://atlas.rsna.org/schemas/2025-11/model.json

Estimation of Cardiac Valve Annuli Motion with Deep Learning

https://arxiv.org/abs/2010.12446v1

v1

eric.kerfoot@kcl.co.uk

Wellcome EPSRC Centre for Medical Engineering at the School of Biomedical Engineering and Imaging Sciences, King’s College London (WT 203148/Z/16/Z).

Data was collected from studies with ethical approval, including an ongoing imaging study for HCM patients and the DETERMINE study for MI patients via the Cardiac Atlas Project.

This model uses a deep learning regression network to identify ten cardiac valve landmarks from long-axis MR images, enabling the quantification of valve annuli motion and morphology for understanding heart function in healthy and pathological states.

The network architecture consists of a sequence of five dilated dense blocks, each applying 2D convolutions with different dilation rates and concatenating results. Each dense block down-samples the spatial dimensions by half. The output volume is then passed to a series of small networks, each with a convolution and a fully-connected layer, to compute one landmark coordinate. All convolution kernels have 3x3 shape.

Provides a fast and precise method for tracking long-axis motion throughout the cardiac cycle, enabling a better understanding of healthy and pathological heart function. It allows for rapid assessment of valve annulus motion and morphology without manual initialization, aiding in the quantification of global strains and markers of systolic/diastolic function.

Diagnosis

Fully automated, requiring no manual initialisation or user input for landmark detection.

Identification of cardiac valve landmarks (mitral, aortic, tricuspid) from long-axis MR images to characterize valve annuli motion and morphology. Applicable for assessing cardiac function and long-axis motion in diverse cohorts, including patients with hypertrophic cardiomyopathy, myocardial infarction, and dilated cardiomyopathy, as well as healthy volunteers.

Unlabeled long-axis cardiac Magnetic Resonance (MR) images (2-chamber, 3-chamber, 4-chamber views).

The network predicts the presence and location of 10 valve landmarks in individual long-axis images. Time-series images are not required, and valve landmarks can be obtained from a single frame. The model can be used as a stand-alone tool or to automatically initialize landmark locations for more robust tracking methods.

The current network does not incorporate temporal consistency, which may be beneficial in cases with poor image quality due to flow artifacts. Post-processing steps may be needed to interpolate landmark positions in such frames. Poor agreement was observed for aortic valve peak strain due to flow artifacts at end-systole.

Recommended for rapid assessment of cardiac valve annulus motion and morphology in large cohorts, particularly for research into healthy and pathological cardiac function. It can serve as a primary tool or as an initialization step for other tracking methods.

Differences between peak long-axis strain measured from predicted landmarks versus manually annotated landmarks were similar to differences observed between multiple expert observers.