Exploring the capabilities of CNNs for 3D angiographic reconstructions from limited projection data using rotational angiography

This study leverages convolutional neural networks (CNNs) to enhance the temporal resolution of 4D angiography for use in intracranial aneurysm (IA) assessment, focusing on the reconstruction of comprehensive volumetric data from truncated projections. Three patient-specific IA geometries were segmented and converted into stereolithography (STL) files which were then used for computational fluid dynamics (CFD) simulations. These simulations modeled blood flow under steady conditions with inlet velocities of 0.25 m/s, 0.35 m/s, and 0.45 m/s. Subsequently, angiograms were simulated by labeling inlet particles to represent contrast bolus injections over durations of 0.5 s, 1.0 s, 1.5 s, and 2.0 s. The angiographic simulations were utilized to emulate a cone-beam CT system performing rotational digital subtraction angiography (DSA), with an exposure time of 10 milliseconds and a temporal resolution of 25 fps. From these simulations, both fully sampled and truncated projection datasets were generated. Fully sampled projection dataset used as ground truth for your CNN, trained to recover the high-fidelity data from truncated reconstructions. The network, designed with ReLU activations, max pooling layers, and upsampling modules, reconstructed volumetric angiographic data. Model performance was assessed using mean squared error (MSE). The trained CNN achieved a MSE of 1×10⁻⁴ across the test dataset, demonstrating high accuracy in reconstructing transient flow patterns and capturing subtle angiographic features. Additionally, we were able to generate angiographic parametric imaging (API) parameters based on the CNN reconstructions, enabling further downstream analyses. Comparative analysis revealed a marked improvement in reconstruction quality compared to traditional FDK methods when using sparse projection datasets. This study highlights the efficacy of CNN-based reconstruction for high-resolution 4D angiography from truncated input data. The approach shows promise for reducing imaging time and radiation exposure, while maintaining diagnostic accuracy. Future work will focus on in vitro validation and real-time implementation to facilitate widespread adoption in neurointerventional workflows.