DeepCA: Deep Learning-Based 3D Coronary Artery Tree Reconstruction from Two 2D Non-Simultaneous X-ray Angiography Projections

Yiying Wang, Abhirup Banerjee, Robin P. Choudhury, Vicente GrauWACV 2025

TL;DR

This paper introduces DeepCA, a novel deep learning framework for reconstructing the 3D structure of coronary arteries from just two standard 2D X-ray angiography (ICA) images. By cleverly simulating motion and using a conditional generative adversarial network, DeepCA can overcome the significant challenge of non-simultaneous image acquisition, paving the way for more accurate and automated cardiac interventions.

What Problem Does This Paper Address?

Cardiovascular diseases (CVDs) are a leading cause of death globally. Invasive X-ray coronary angiography (ICA) is a crucial diagnostic tool, but it typically captures only two 2D projections. This limited view makes it difficult for clinicians to fully understand the complex 3D anatomy of coronary arteries, especially when dealing with vessel overlap, foreshortening, and motion artifacts caused by breathing and heartbeats. While more projections can provide better clarity, they increase risks associated with contrast agents and radiation. Existing methods often require extensive manual intervention and struggle to correct for the non-rigid motion that occurs between non-simultaneous projections, leading to inaccurate 3D reconstructions. This paper aims to bridge this gap by developing a deep learning approach that can accurately reconstruct 3D coronary trees from these challenging two-view, non-simultaneous ICA images.

Methods

The proposed DeepCA pipeline consists of two main blocks:

  1. Data Preprocessing Block: This block is designed to generate realistic training data. It starts with 3D coronary artery data from Computed Tomography Angiography (CCTA). To simulate the conditions of real ICA, it introduces rigid transformations to the CCTA data before generating two 2D projections. Crucially, it simulates the non-rigid motion that occurs between non-simultaneous ICA acquisitions. The resulting 2D projections are then binarized and backprojected to create a 3D input for the reconstruction network.
  2. 3D Reconstruction with Motion Compensation Block: This is the core of DeepCA. It employs a Wasserstein conditional generative adversarial network (WCGAN) with gradient penalty. This architecture allows the model to learn to generate a desired 3D reconstruction (the output) conditioned on the noisy, motion-affected 3D backprojection input.
  3. Generator (G): A 3D U-Net architecture is used as the backbone, enhanced with novel latent convolutional transformer layers (CTLs). These CTLs are designed to capture long-range dependencies and relationships within the feature maps in the latent space, which is crucial for understanding the global structure of the coronary tree.
  4. Critic (D): A dynamic snake convolutional (DSConv) critic is introduced. This critic is specifically designed to handle the quasi-tubular structures of coronary arteries. It uses a novel kernel that can adapt its shape to better extract both global tubular and local features, improving the ability to distinguish vessel structures.
  5. Loss Function: The WCGAN-GP objective is used to ensure stable training and generate realistic reconstructions. An additional L1 loss is incorporated to enforce alignment with the ground truth.

The key novelty lies in the combination of simulating motion for training, the use of latent transformers for better feature representation, and the specialized DSConv critic for tubular structures, all within a conditional GAN framework to implicitly compensate for motion.

Datasets

  • CCTA Dataset: A public dataset of 3D binary segmented coronary trees was used for training and validation. This dataset was split into 75% for training, 15% for validation, and 10% for testing. The Right Coronary Artery (RCA) was specifically focused on.
  • Real ICA Dataset: A clinical dataset from 8 patients was collected for evaluation. This dataset consisted of real, non-simultaneous ICA projections. For two patients, three projections were available, while for the rest, two were used. This dataset is considered an "out-of-distribution" domain, testing the model's generalization capabilities.

Key Insights and Ideas

  • Simulating Motion is Key for Generalization: The paper highlights that training on simulated projections that mimic real-world motion artifacts is crucial for achieving good performance on actual, unseen ICA data. This is a significant step beyond previous methods that relied solely on synthetic or perfectly aligned data.
  • Latent Transformers for Global Structure: The introduction of latent convolutional transformer layers in the generator allows the model to learn global relationships between different parts of the coronary tree, which is essential for accurate 3D reconstruction, especially when dealing with limited views.
  • Specialized Critic for Tubular Structures: The dynamic snake convolutional critic is a clever design choice. Coronary arteries are inherently tubular, and a critic that can specifically adapt to these shapes is likely to be more effective at distinguishing them from background noise and artifacts.
  • Implicit Motion Compensation: The WCGAN framework, by learning to map noisy inputs to clean outputs, implicitly learns to compensate for the non-rigid motion present in the input projections, a major challenge in this field.

Key Results

  • Superior Performance on CCTA Data: DeepCA outperformed four baseline models on the CCTA test dataset across various metrics, including Ot(d) (overlap using sweeping distance threshold) and CD(l2) (Chamfer L2 distance). For example, it achieved an Ot(1) of 64.63% and an Ot(2) of 76.25% on the CCTA test set.
  • Excellent Generalization to Real ICA Data: The model demonstrated significant improvements on the real clinical ICA dataset, particularly for the second and additional projection planes. It showed substantial gains in Ot(1) (38.57% improvement) and CD(l2) (42.25% improvement) compared to the best baseline models for these challenging projections.
  • Recovery of Missing Features: Qualitative results showed DeepCA's ability to recover missing vascular sections in the original ICA data, likely due to acquisition errors or contrast agent timing.
  • Ablation Study Confirms Component Importance: An ablation study showed that combining the WCGAN-GP, latent CTLs, and DSConv critic yielded the best results, indicating the importance of each component.

Limitations and Future Work

  • Focus on RCA: The study primarily focused on the Right Coronary Artery (RCA). Future work could extend this to other coronary arteries.
  • Motion Artifacts: While DeepCA implicitly compensates for motion, severe motion artifacts can still lead to incomplete reconstructions, as observed in some qualitative examples.
  • Potential for Automation: The authors suggest that combining DeepCA with automated coronary vessel segmentation could enable end-to-end automated 3D reconstruction for real-time cardiac interventions. This points towards future research in integrating segmentation and reconstruction pipelines.
  • Further Evaluation: While the paper introduces a new evaluation metric, further validation on larger and more diverse clinical datasets would be beneficial.

Summary updated

BibTeX
@inproceedings{wang2025deepca,
  author = {Yiying Wang and Abhirup Banerjee and Robin P. Choudhury and Vicente Grau},
  title = {{DeepCA: Deep Learning-Based 3D Coronary Artery Tree Reconstruction from Two 2D Non-Simultaneous X-ray Angiography Projections}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure