Abstract
Feature tracking Cardiac Magnetic Resonance (CMR) has recently emerged as an area of interest for quantification of regional cardiac function from balanced, steady state free precession (SSFP) cine sequences. However, currently available techniques lack full automation, limiting reproducibility. We propose a fully automated technique whereby a CMR image sequence is first segmented with a deep, fully convolutional neural network (CNN) architecture, and quadratic basis splines are fitted simultaneously across all cardiac frames using least squares optimization. Experiments are performed using data from 42 patients with hypertrophic cardiomyopathy (HCM) and 21 healthy control subjects. In terms of segmentation, we compared state-of-the-art CNN frameworks, U-Net and dilated convolution architectures, with and without temporal context, using cross validation with three folds. Performance relative to expert manual segmentation was similar across all networks: pixel accuracy was ~97%, intersection-over-union (IoU) across all classes was ~87%, and IoU across foreground classes only was ~85%. Endocardial left ventricular circumferential strain calculated from the proposed pipeline was significantly different in control and disease subjects (-25.3% vs -29.1%, p = 0.006), in agreement with the current clinical literature.
Introduction
Quantifying regional heart function through "strain"—the deformation of heart muscle during a cardiac cycle—is critical for diagnosing and monitoring cardiovascular diseases. While "tagged" MRI remains the gold standard, its slow acquisition and complex analysis limit clinical utility. Feature Tracking (FT) offers a faster alternative by tracking anatomical landmarks in standard "cine" MRI sequences. However, traditional FT methods suffer from two primary drawbacks: the requirement for manual contour initialization, which introduces inter-observer variability, and "tracking drift" caused by pairwise frame-to-frame motion estimation.
Methods
The proposed pipeline automates both the identification and tracking of cardiac structures through a two-stage process.
Automated Segmentation
The system employs Deep Convolutional Neural Networks (CNNs) to segment the Left Ventricle (LV) myocardium, the LV blood pool, and the Right Ventricle (RV) blood pool. The researchers evaluated U-Net and Dilated Convolution (DC) architectures. A key finding was the benefit of "Temporal Context," where providing the network with both the current frame and the End Diastole (relaxed state) frame significantly improved segmentation consistency.
Joint Spline Optimization
To address tracking drift, the authors replaced pairwise tracking with a global optimization approach. Quadratic Bézier splines are fitted to the cardiac boundaries and optimized across all frames of the cycle simultaneously.
This joint optimization includes physics-informed regularization penalties for unrealistic motion and extreme curvature, ensuring the resulting deformation is biologically plausible. By treating the entire heartbeat as a single problem, the model avoids the accumulation of errors inherent in sequential methods.
Results and Clinical Validation
The study utilized a dataset of 63 subjects, including 42 patients with hypertrophic cardiomyopathy (HCM) and 21 healthy controls. The segmentation model achieved high fidelity, with a pixel accuracy of ~97% and an IoU of ~87% compared to expert manual tracings.
Crucially, the pipeline demonstrated clinical relevance by successfully measuring endocardial LV circumferential strain. The system identified a statistically significant difference between healthy subjects (-25.3%) and HCM patients (-29.1%, $p=0.006$), aligning with established clinical literature. The model proved robust even in cases of severe hypertrophy where cardiac anatomy is significantly distorted.
Discussion
The deep learning approach effectively handles challenging features like papillary muscles, which often confound simpler automated algorithms. Furthermore, the system is highly efficient, with segmentation running at 75 frames per second and the entire tracking process completing in approximately 2 seconds per sequence. This speed and automation make it suitable for real-world clinical environments.
While the current model is highly effective, it faces challenges such as the "aperture problem"—difficulty tracking motion along smooth edges without distinct features—and sensitivity to non-uniform image illumination. Future developments will focus on integrating CNN-derived features directly into the spline optimization to enhance tracking in feature-sparse regions and using data augmentation to improve robustness to lighting variations.
Summary updated