Segmentation of the Left Atrium in Cardiovascular Magnetic Resonance Images of Patients with Myocarditis

Summary

The three-dimensional fully convolutional network-enhanced cine analysis enables precise left atrial function assessment in myocarditis, improving early systolic-diastolic dysfunction detection and reducing ejection fraction prediction errors for clinical diagnosis and monitoring of atrial mechanical dysfunction.

Abstract

Cardiovascular magnetic resonance (CMR) cine sequences serve as the cornerstone imaging technique for evaluating dynamic left atrial (LA) function in myocarditis patients. By capturing three-dimensional motion characteristics throughout the cardiac cycle with high temporal resolution, this modality provides critical data for analyzing myocardial contractile coordination and wall motion abnormalities. Key technological innovations, such as dynamic modeling and strain-encoded imaging, enable quantitative assessment of early-stage LA systolic-diastolic dysfunction in myocarditis. However, the primary challenges in cine sequence segmentation involve dynamic artifacts and spatiotemporal continuity modeling of thin-walled structures. Traditional threshold-based segmentation methods demonstrate limited consistency in dynamic sequences due to their inability to capture motion patterns. Deep learning approaches utilizing three-dimensional fully convolutional network (3D-FCN) achieved superior accuracy through three strategic enhancements: (1) Spatiotemporal feature fusion: This employed 3D convolutional kernels to simultaneously extract spatial structures and temporal dimensional features, thereby reducing motion blurring effects. (2) Dynamic skip connections: Incorporated within encoder-decoder architectures, these connections strengthened deformation correlation modeling across different cardiac phases through cross-temporal feature propagation. (3) Lightweight design: By utilizing patch-wise processing and depthwise separable convolutions, computational efficiency was optimized for real-time processing of large-scale four-dimensional datasets. The 3D-FCN achieved a Dice coefficient of 0.921 for LA segmentation, representing a 12.3% improvement over conventional methods. This design reduced the LA ejection fraction prediction error from 8.7% to 3.2%. The segmentation results directly facilitated the calculation of quantitative metrics, including LA volume-time curves and strain rates. These metrics supported the clinical diagnosis of myocarditis-associated atrial mechanical dysfunction.

Introduction

Cardiovascular magnetic resonance (CMR) cine sequences, as the core imaging modality for evaluating dynamic left atrial (LA) function in myocarditis patients, provide critical data for identifying abnormalities in myocardial contractile coordination and early diastolic dysfunction, leveraging their high temporal resolution and three-dimensional motion capture capabilities^1,2,3,4,5. Through strain-encoded imaging and dynamic motion modeling techniques, CMR cine sequences enable precise detection of myocardial edema and fibrosis, which correlate strongly with histopathological changes and predict major adverse cardiovascular events^6,7,8,9. CMR feature tracking enables quantitative assessment of LA strain, reflecting its reservoir, conduit, and pump functions. Elevated left atrial pressure, which is associated with left ventricular diastolic dysfunction, can be indirectly evaluated through LA strain rate analysis via CMR, providing insights into abnormal left ventricular filling patterns^8,10,11,12.

In myocarditis, specific strain metrics -- such as passive strain and active strain -- exhibit strong correlations with myocardial fibrosis, allowing early detection of compensatory dysfunction in LA mechanics^13,14,15,16. Notably, reductions in LA strain often precede structural dilation in myocarditis patients, and CMR cine sequences sensitively capture these early remodeling patterns, offering a critical window for timely clinical intervention. However, cine sequence segmentation faces multiple challenges: the thin-walled LA structure is susceptible to respiratory motion artifacts, leading to errors in ejection fraction calculation, while traditional threshold-based segmentation methods systematically overestimate the LA minimal volume due to their inability to capture asymmetric deformation patterns^15,17.

In recent years, compressed sensing technology has reduced scan time via k-space under-sampling, demonstrating high correlation with gold-standard measurements for LA passive ejection fraction, though underestimation of total ejection fraction and active pump function remains to be optimized^18,19,20,21. Notably, deep learning-driven cardiac motion prediction models have achieved accurate dynamic LA functional modeling in free-breathing sequences by integrating phase velocity encoding and adaptive motion correction algorithms^22,23. This study aims to use deep learning algorithms to establish a critical foundation for clinical diagnosis of myocarditis-associated LA mechanical dysfunction, enabling personalized risk stratification and prognostic prediction for myocarditis patients.

Protocol

The Ethics Committee of Chengdu Medical College determined that this study met the criteria for waiver of informed consent. The research adhered to the principles outlined in the Declaration of Helsinki.

1. Patient selection

1. Use the following inclusion criteria: patients with myocarditis meeting clinical criteria for acute myocarditis, including post-viral symptoms, elevated biomarkers, and electrocardiogram (ECG) abnormalities.
2. Exclude patients with the following: ischemic, valvular, or congenital heart disease; secondary myocardial injury from drugs, chemotherapy or infections; non-diagnostic imaging; high-risk conditions (pregnancy, renal dysfunction or hemodynamic instability).

2. CMR imaging protocol and scanning parameters

1. Scanner configuration
   1. Operate a 3.0 Tesla magnetic resonance imaging scanner equipped with a 32-channel cardiac phased-array coil within 14-28 days post-symptom onset.
   2. Execute coil verification using the console's Coil Check function, followed by an auto-calibration routine.
      NOTE: Patient positioning requires supine orientation with the coil centered at the left mid-axillary line.
2. Left atrial functional assessment
   1. Initiate balanced steady-state free precession cine imaging through Protocol Manager > Cardiac > Cine_balanced steady-state free precession (bSSFP) pathway.
   2. Configure retrospective electrocardiogram gating with the Trigger Window parameter set to 15% for arrhythmia accommodation.
   3. Achieve full cardiac cycle coverage by programming 33 ms temporal resolution (30 phases) via 12 k-space segments.
      NOTE: Key parameters include repetition time/echo time 3.0/1.5 ms, flip angle 45 °, bandwidth 930 Hz/pixel, and matrix 256×256.
3. Standard cardiac plane acquisition
   1. Obtain scout images using a three-plane localizer sequence, taking approximately 18 s.
   2. Prescribe four-chamber view by aligning the imaging plane through the left ventricular apex and mitral valve center on axial localizer.
   3. Generate a two-chamber view by duplicating the four-chamber prescription and applying 60°clockwise rotation.
4. Parallel imaging implementation
   1. Activate generalized autocalibrating partially parallel acquisition using the console's Acceleration tab with parallel imaging factor 2.
   2. Enable partial Fourier sampling by setting Phase Partial Fourier to a 6/8 configuration.
   3. Perform reference calibration with 24 autocalibration signal lines requiring 22 s.
      NOTE: Key scan parameters include TR/TE = 3.0 ms/1.5 ms, flip angle = 45°, and bandwidth = 930 Hz/pixel.
5. Left atrium-specific imaging
   1. Identify anatomical landmarks, including the mitral annulus at mid-diastole, the funnel-shaped left atrial roof 15 mm from the septum, and the left atrial appendage (LAA) orifice demonstrating distinct lobar architecture.
   2. Align long-axis planesperpendicular to the four-chamber view spanning from pulmonary vein orifices to left atrial appendage tip.
      NOTE: Three essential anatomical landmarks guide CMR plane localization: the mitral annulus, requiring mid-diastolic visualization of both leaflets, the funnel-shaped LA roof centered 15 mm from the septum at the right superior pulmonary vein ostium, and the LAA orifice demonstrating clear lobar architecture with distinct borders from the left superior pulmonary vein.
6. Short-axis stack acquisition
   1. Prescribe contiguous short-axis slices parallel to the mitral annulus plane using diastolic two-chamber view guidance.
   2. Configure slice geometry with 3 mm thickness, totaling 15 slices covering from the pulmonary vein origins to the left ventricular apex.
7. High-resolution imaging
   1. Employ a zoomed balanced steady-state free precession sequence achieving 0.8×0.8×2.5 mm resolution with a bandwidth of 1200 Hz/pixel.
   2. Position diastolic saturation band over left atrial appendage outflow at 50 ms post-R-wave using 10 mm thickness.
8. Respiratory motion control
   1. Deploy navigator gating by activating the diaphragm tracking function on the right hemidiaphragm dome with a 5 mm acceptance window.
   2. Minimize breath-hold duration by enabling compressed sensing reconstruction at an acceleration factor of 4.0 and implementing real-time steady-state free precession sequences.
      NOTE: To ensure patient safety, this study strictly adhered to the Society for cardiovascular magnetic resonance (SCMR) position paper (2020) on clinical indications for cardiovascular magnetic resonance²⁴.

3. Data preprocessing and quality control

1. Utilize non-rigid B-spline registration to address respiratory and arrhythmia-induced misalignment (maximal tolerance ± 4 mm) for motion correction.
2. Employ N4 bias field correction and histogram matching to standardize intensity across scanners for signal homogenization.
3. Perform gold-standard annotations by two expert radiologists (>5 years' experience) in 3D Slicer according to the guideline²⁵.
   NOTE: System requirements, installation steps, basic functional modules, and workflow are based on the official guide²⁶.
4. Resolve thin LA walls (<2 mm) using 50% gray-scale gradient thresholds. Inter-observer agreement for annotations exceeded ICC > 0.91 and Dice > 0.95, ensuring robust ground truth for model training and validation.
   NOTE: All data undergoes reversible anonymization, which protects privacy through adversarial perturbations and visible watermarks while allowing authorized decryption and restoration. Temporarily stored data is encrypted with strict access controls and is cryptographically erased upon disposal²⁷.

4. Spatiotemporal feature fusion technology

1. Employ a deep learning approach based on a 3D fully convolutional network (3D-FCN) to enhance left atrial segmentation accuracy across all cardiac phases.
2. Integrate spatiotemporal feature pyramid modules incorporating dilation rates of 1, 2, 4, and 8 within the 3D-FCN architecture to capture myocardial motion patterns across multiple resolution scales.
   NOTE: Feature extraction employs five encoder stages with channel dimensions [32, 64, 128, 256, 512] progressively increasing with each downsampling operation.

5. Network architecture

1. Multiscale spatiotemporal modeling
   1. Deploy hierarchical 3D convolutional kernels (kernel size 3 × 3 × 3, stride 1 × 1 × 1, dilation rate 2) at three resolutions (16 × 16 × 16, 32 × 32 × 32, 64 × 64 × 64).
   2. Capture cross-scale spatiotemporal correlations covering local fiber motion to global ventricular contraction.
2. Input configuration
   1. Use a 4D tensor input (channels × time × height × width = 4 × 30 × 256 × 256) covering a full cardiac cycle.
   2. Include four-chamber and two-chamber cine sequences with 30 frames per cycle at 33 ms temporal resolution.

6. Dynamic skip connection

NOTE: A cross-cycle dynamic skip connection is designed with adaptive weight allocation to address the temporal characteristics of cine sequences.

1. Encoder
   1. Design five-level 3D residual blocks (each containing four 3 × 3 × 3 convolutions, with channel count expanding from 64 to 512) to extract spatiotemporal features.
   2. Dynamically suppress respiratory motion artifacts and blood flow noise by computing spatiotemporal weight matrices in a spatiotemporal attention gate.
   3. Extract multiscale spatiotemporal features, with output tensor dimensions [C × T × H × W] = [64 × 30 × 256 × 256, 128 × 30 × 128 × 128, 256 × 30 × 64 × 64, 512 × 30 × 32 × 32, 1024 × 30 × 16 × 16] (T = 30 frames/cycle) in five-layer pyramid.
2. Decoder
   1. Combine high-level semantic features (64 × 64 × 64 resolution) with low-level resolution features (16 × 16 × 16) via 3D transposed convolution (kernel size 2 × 2 × 2, stride 2) for cross-scale feature fusion.
   2. Ensure synchronized spatiotemporal refinement in temporal interpolation (cubic spline interpolation, temporal factor 0.5).
   3. Establish temporal dependencies in bidirectional gated recurrent unit modules by leveraging both forward and backward hidden states to capture long-range contextual patterns while mitigating vanishing gradient issues through gating mechanisms.
   4. Dynamically fuse features from adjacent 5 frames in temporal alignment convolution (kernel size 3 × 3 × 3, stride 1 × 1 × 1, dilation rate 2).

7. Adaptive weight allocation strategy

1. Configure dual-branch processing pathways within the spatiotemporal attention gating module to dynamically balance spatial and temporal feature contributions during network operations.
2. Compute spatial weight matrices for 3 × 3 deformable convolutions, emphasizing myocardial boundaries (thin-walled LA structures <2 mm) and compensating for respiratory motion artifacts in the spatial attention branch.
3. Prioritize critical cardiac phase correlations by generating temporal weighting coefficients based on tissue deformation metrics derived from myocardial displacement vectors.

8. Residual pathway stabilization

1. Randomly omit temporally dropped connections during training to enhance robustness for preventing gradient explosion and feature collapse.
2. Constrain weight matrices to stabilize feature propagation.
   1. Apply L2 regularization with a coefficient of zero point zero one to constrain weight matrix updates.
   2. Limit the maximum weight gradient value to one point zero for clipping.
   3. Initialize weights using a scaling factor of the square root of two divided by the number of input nodes.
3. Channel reduction with depthwise separable convolution
   1. Initialize depthwise convolution using kernel size 3 in all three dimensions with group count matching the input channel dimension.
   2. Apply pointwise convolution with 1×1×1 kernel configuration to project feature channels from 512 to 64 dimensions.
   3. Execute batch normalization and parametric rectified linear unit activation before feature concatenation operations.
4. Segmentation synthesis, and artifact suppression
   1. Fuse spatiotemporal modeling output with adaptive attention feature maps through channel-wise concatenation.
   2. Enhance myocardial boundary delineation using multiscale contour constraints applied specifically to regions thinner than 2 mm.
   3. Implement motion artifact suppression through temporally-weighted feature stabilization across cardiac phases.
   4. Generate final left atrial segmentation via 1×1×1 convolutional output layer with sigmoid activation.
      NOTE: For visual validation, the segmentation results are displayed using 3D surface rendering overlaid on original image slices (semi-transparent mask overlay technique) to ensure anatomical continuity of the left atrial thin-wall structures. For quantitative evaluation, metrics including Dice coefficient and Hausdorff distance (HD) are employed.

9. Other troubleshooting guidance

1. Adjust the ECG gating parameters to achieve 25-35 ms temporal resolution for optimal cardiac phase capture.
2. Implement respiratory compensation using 5 mm navigator gating windows to minimize breathing-related motion artifacts.
3. Ensure proper patient positioning and provide clear breathing instructions throughout the imaging procedure.

Results

On a CMR dataset comprising 200 myocarditis patients, the proposed framework achieved a Dice coefficient of 0.921 for LA segmentation, representing a 12.3% improvement over baseline models, with a processing speed of 18 frames per second (fps). By employing dynamic kernel distillation to transfer high-level abstract features to low-level features, the HD for LA minimum volume (LAV_min) was optimized from 3.2 mm to 1.7 mm (reduction: 46.9%). For thin-walled LA structures (thickness <2 mm), HD decreased from 4...

Discussion

This study achieved precise quantification of LA structure and function in myocarditis patients through standardized CMR protocols and dedicated LA scanning parameters, combined with efficient feature extraction and automated rapid segmentation using 3D-FCN. The synergistic interaction between optimized imaging parameters and algorithmic innovations establishes a robust technical foundation for clinical translation.

CMR remains the cornerstone tool for assessing LA structural and functional ab...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
3.0 T MRI scanner	Siemens Healthineers	Skyra 3.0 T	High-resolution clinical/research MRI system with advanced cardiac and body imaging capabilities.
3D Slicer	Open-source community	https://www.slicer.org/	Free, open-source software for medical image analysis (segmentation, registration, 3D visualization). Supported by NIH.
PyTorch	Meta Platforms, Inc.	https://pytorch.org/	Open-source deep learning framework with dynamic computation graphs, widely used for AI research and model deployment. Supports GPU acceleration.