Iterated Residual Graph Convolutional Neural Network for Personalized Three-Dimensional Reconstruction of Left Myocardium from Cardiac MR Images

Xuchu Wang, Yue Yuan, Minghua Liu, Yanmin Niu2023 · DOI

TL;DR

This paper introduces a novel deep learning approach for creating personalized 3D models of the left ventricle from cardiac MRI scans. By using an iterated residual graph convolutional neural network, the method refines an initial mesh to accurately capture individual cardiac anatomy, offering a more robust and detailed reconstruction than existing methods.

What Problem Does This Paper Address?

Accurate 3D reconstruction of the left myocardium (heart muscle) is crucial for diagnosing and treating cardiac diseases. Existing methods often struggle with the inherent variability in individual heart shapes and the dynamic nature of the heart. Traditional 3D modeling requires expert knowledge, while automated reconstruction methods can be limited by the similarity to average models or the complexity of processing raw image data. Deep learning has shown promise, but existing approaches often fall into voxel-based, point-cloud-based, or triangular-mesh-based categories, each with limitations in accuracy, storage, or the ability to represent fine surface details. This paper aims to overcome these challenges by developing a personalized, accurate, and robust 3D reconstruction method for the left myocardium.

Methods

The proposed method employs a multi-stage approach:

  • Initial Mesh Generation: A UNet++ based segmentation algorithm is used to process stacked 2D cardiac MR slices, followed by the Cubify algorithm to generate an initial triangular mesh. This mesh is then post-processed to simplify facets.
  • Vertex Feature Learning: To address the limitations of the initial mesh (sharp edges, corners), a vertex feature learning module is introduced. This module uses an encoder-decoder neural network to extract deep features from the initial mesh's point cloud, capturing the skeleton and local relationships of the left myocardium at various receptive fields. This is achieved through a hierarchical point set abstraction process involving sampling, grouping, and a mini-PointNet layer.
  • Iterated Mesh Deformation: The core of the method is an iterated residual graph convolutional neural network (ResGCNet). This network gradually refines the initial triangular mesh. In each iteration, the learned vertex features are fused with the mesh's current vertex features to guide the deformation. This iterative process, combined with residual connections to prevent gradient explosion, allows for progressive refinement and improved accuracy. The loss function incorporates Chamfer distance, normal vector distance, and shape regularization to ensure similarity to the ground truth, surface smoothness, and prevention of mesh degradation.

Datasets

The study utilized a dataset of cardiac MR volumes from 20 subjects obtained from real clinical institutions. The acquisition equipment included Siemens and Philips MRI scanners. For each subject, the left ventricular myocardium tissue was manually labeled by two trained students and cross-validated by a clinical expert. The dataset was augmented and split into training (60 cases), validation (20 cases), and testing (10 cases) sets.

Key Insights and Ideas

  • Personalization is Key: The paper emphasizes that individual anatomical variations are a major hurdle in cardiac reconstruction, and their method directly addresses this by starting with an individual's segmentation.
  • Mesh Deformation Guided by Deep Features: Instead of relying solely on geometric properties, the method leverages deep learned features from the point cloud to guide the mesh deformation, leading to more accurate and robust results.
  • Iterative Refinement for Accuracy: The iterated nature of the ResGCNet allows for gradual improvement of the mesh, correcting initial inaccuracies and smoothing the surface over multiple steps.
  • Balancing Detail and Smoothness: The combination of Chamfer distance, normal vector distance, and shape regularization in the loss function helps achieve a balance between capturing fine geometric details and maintaining a smooth, realistic surface.

Key Results

  • Quantitative Performance: The proposed method achieved the smallest Chamfer distance among the compared methods (MC, Poisson, HRBF), indicating superior similarity to the ground truth. It also performed competitively on other metrics like Normal score and F1 scores.
  • Qualitative Performance: Visualizations show that the proposed method produces smoother and more geometrically accurate reconstructions compared to MC, Poisson, and HRBF, particularly in preserving local geometric structures. The iterative deformation process clearly demonstrates the refinement from a coarse initial mesh to a detailed final surface.
  • Computational Efficiency: The proposed method achieved a reconstruction speed of 2614 ms per subject, significantly faster than traditional methods like HRBF (approx. 59,000 ms) and competitive with MC and Poisson.

Limitations and Future Work

  • Segmentation Dependency: The quality of the final reconstruction is directly dependent on the accuracy of the initial left myocardial segmentation.
  • Uneven Point Cloud Density: The vertex feature learning module, similar to PointNet++, can be affected by uneven density distribution in the point cloud, potentially leading to loss of important features.
  • Future Directions: The authors suggest introducing an adaptive density learning strategy to improve feature extraction. They also propose integrating the 3D segmentation and reconstruction steps into a single, unified model to further enhance performance.

Summary updated

BibTeX
@misc{wang2023iterated,
  author = {Xuchu Wang and Yue Yuan and Minghua Liu and Yanmin Niu},
  title = {{Iterated Residual Graph Convolutional Neural Network for Personalized Three-Dimensional Reconstruction of Left Myocardium from Cardiac MR Images}},
  year = {2023},
  doi = {10.3390/s23177430},
}
Figure