Preserving Cardiac Integrity: A Topology-Infused Approach to Whole Heart Segmentation

Guang Yang, Ieee Senior Member, Chenyu Zhang, Wenxue Guan, Xiaodan Xing, Guang YangCARE @ MICCAI 2024 · DOI

TL;DR

This paper introduces a novel "topology-preserving module" integrated into deep neural networks for whole-heart segmentation. By explicitly encoding anatomical constraints like containment and exclusion, their method significantly improves segmentation accuracy and robustness, particularly for challenging MRI data, while maintaining computational efficiency.

What Problem Does This Paper Address?

Accurate segmentation of the whole heart is crucial for diagnosing and monitoring cardiovascular diseases (CVDs). However, this task is fraught with challenges: the heart's shape changes dramatically throughout the cardiac cycle, clinical imaging can suffer from motion artifacts and poor contrast-to-noise ratios, and data from different centers or scanners (domain shift) can lead to inconsistent results. While deep learning methods have shown promise, they often struggle to capture global structural relationships, such as the fact that the left ventricle myocardium always encloses the left ventricle. This paper aims to bridge this gap by developing a method that not only segments the heart accurately but also preserves its inherent anatomical topology.

Methods

The core innovation is a topology-preserving module (TPM) that is integrated into a standard 3D U-Net-like architecture. The TPM works by identifying "key voxels" that violate predefined topological constraints between different heart structures. These constraints are based on two fundamental relationships:

  • Containment: One structure is entirely within another (e.g., Myocardium encloses Left Ventricle).
  • Exclusion: Two structures cannot be adjacent (e.g., Right Atrium and Ascending Aorta are distinct).

The TPM encodes these relationships using 3D convolutions and a novel loss function. Specifically, it identifies pairs of neighboring voxels that violate these constraints. These violations are then penalized through an additional loss term, Ltp, which is added to the standard cross-entropy (LCE) and Dice loss (Ldice). This end-to-end training approach forces the network to learn not just pixel-wise accuracy but also to respect the global anatomical structure, correcting topological errors that might arise from standard segmentation approaches.

Datasets

The authors evaluated their method on the WHS++ challenge dataset. This dataset comprises 104 CT and 106 MRI volumes, each with manual segmentations of seven whole-heart substructures. The data is noted to be acquired from various clinics with different scanners, leading to variations in image quality, resolution, and voxel spacing, making it a realistic and challenging benchmark.

Key Insights and Ideas

  • Explicitly encoding topology matters: The paper demonstrates that simply optimizing for pixel-wise accuracy (like Dice score) is insufficient for robust anatomical segmentation. Incorporating explicit topological constraints significantly improves the quality and anatomical plausibility of the segmentation.
  • 3D convolutions are key for spatial relationships: The use of 3D convolutions within the TPM is crucial for effectively capturing and encoding the spatial relationships between neighboring voxels in the 3D cardiac volume.
  • Topology preservation is particularly beneficial for MRI: While the method improves performance on both CT and MRI, the results suggest a more pronounced benefit for MRI data, which is often more challenging due to lower image quality and greater variability.
  • Correcting "impossible" segmentations: The method can fix topological errors that even strong baselines like nnUNet might produce, such as the left ventricle not being fully enclosed by its myocardium or incorrect separation of adjacent structures.

Key Results

The proposed topology-infused approach achieved state-of-the-art results on the WHS++ dataset.

  • CT Data: The method achieved a Dice coefficient of 0.949 ± 0.035, outperforming baselines like UNet, 3D-UNet, and nnUNet.
  • MRI Data: The method achieved a Dice coefficient of 0.897 ± 0.023, also showing significant improvements, particularly in reducing Hausdorff Distance (HD) to 10.001 ± 4.671.
  • Qualitative Improvements: Visual results (Figure 4) show that the proposed method produces segmentations with fewer topological violations, correctly enforcing relationships like the LV being surrounded by Myo and improving the continuity of structures like the Pulmonary Artery (PA).

Limitations and Future Work

  • Manual Annotation Ambiguity: The authors acknowledge that manual annotations, especially for MRI, can have some ambiguity due to large voxels or subjective interpretations. This can limit how perfectly the method can adhere to the ground truth in all cases.
  • Generalizability to Other Organs: While successful for the heart, the applicability of this topology-preserving module to other organs with different topological constraints would be an interesting avenue for future research.
  • Computational Cost: Although the method is described as efficient, the use of 3D convolutions inherently increases computational demands compared to 2D methods. Further optimization for speed might be explored.

Summary updated

BibTeX
@inproceedings{member2024preserving,
  author = {Guang Yang, IEEE Senior Member and Chenyu Zhang and Wenxue Guan and Xiaodan Xing and Guang Yang},
  title = {{Preserving Cardiac Integrity: A Topology-Infused Approach to Whole Heart Segmentation}},
  booktitle = {CARE @ MICCAI},
  year = {2024},
  doi = {10.1007/978-3-031-87009-5\_3},
}
Figure