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

TL;DR

DeepCA presents a novel deep learning pipeline for 3D coronary tree reconstruction from two non-simultaneous ICA projections by learning motion-robust mappings from simulated CCTA-derived projections. The architecture combines a Wasserstein conditional GAN with gradient penalty, latent convolutional transformer layers, and a dynamic snake convolutional critic to implicitly compensate for non-rigid motion between projections. By training on simulated projections from CCTA and evaluating on a real ICA dataset, DeepCA demonstrates strong vessel topology preservation and recovery of missing features, with notable generalisation to unseen data. This approach enables end-to-end automated reconstruction potential for real-time guidance during cardiac interventions, addressing a critical need in limited-view coronary imaging.

Abstract

Cardiovascular diseases (CVDs) are the most common cause of death worldwide. Invasive x-ray coronary angiography (ICA) is one of the most important imaging modalities for the diagnosis of CVDs. ICA typically acquires only two 2D projections, which makes the 3D geometry of coronary vessels difficult to interpret, thus requiring 3D coronary artery tree reconstruction from two projections. State-of-the-art approaches require significant manual interactions and cannot correct the non-rigid cardiac and respiratory motions between non-simultaneous projections. In this study, we propose a novel deep learning pipeline named \emph{DeepCA}. We leverage the Wasserstein conditional generative adversarial network with gradient penalty, latent convolutional transformer layers, and a dynamic snake convolutional critic to implicitly compensate for the non-rigid motion and provide 3D coronary artery tree reconstruction. Through simulating projections from coronary computed tomography angiography (CCTA), we achieve the generalisation of 3D coronary tree reconstruction on real non-simultaneous ICA projections. We incorporate an application-specific evaluation metric to validate our proposed model on both a CCTA dataset and a real ICA dataset, together with Chamfer $\ell_2$ distance. The results demonstrate promising performance of our DeepCA model in vessel topology preservation, recovery of missing features, and generalisation ability to real ICA data. To the best of our knowledge, this is the first study that leverages deep learning to achieve 3D coronary tree reconstruction from two real non-simultaneous x-ray angiographic projections.

Figures (13)

• Figure 1: The overall workflow of our proposed DeepCA pipeline consists of a data preprocessing block and a 3D reconstruction with motion compensation block. (a) The data preprocessing block generates two simulated ICA projections from 3D CCTA data, including simulated motion between projections, and then produces the 3D model input via performing backprojection on the two simulated projections. (b) The 3D reconstruction with motion compensation block receives the 3D backprojection input to train a deep neural network for 3D coronary tree reconstruction learned from the CCTA data, implicitly compensating for any motion.
• Figure 2: The proposed DeepCA model architecture includes a conditional generator and a critic. The conditional generator is based on 3D U-Net with additional proposed convolutional transformer layers in the latent space. The generator produces corresponding reconstructed results according to the input condition. The latent convolutional transformers are built on convolutional embeddings following 8 transformer layers. The predicted results and the corresponding ground truth are concatenated with the input separately, which are then sent to the critic. The proposed critic uses both dynamic snake convolution and traditional convolution at the first layer to extract both global tubular and local features, and then applies several downsamplings to generate the critic loss.
• Figure 3: The dynamic snake convolution (DSConv). For each voxel (the yellow voxel in the figure as an example) in the feature map, the DSConv flattens the whole kernel along different axes with random offsets to extract different dynamic feature maps for X-, Y-, and Z-axes separately and then concatenates these feature maps together as the convolution output.
• Figure 4: Three 3D reconstruction results on the CCTA test dataset by our DeepCA model. From top to bottom: three CCTA samples. From left to right: predicted reconstruction, ground truth, and the corresponding voxel-wise prediction errors in terms of CD($\ell_2$).
• Figure 5: An example of 3D reconstruction for the RCA branch of a patient. Left: the original ICA data. Right: 3D reconstruction by our proposed DeepCA model.