Live reconstruction in ultra-fast dynamic computed tomography

Ultra-fast dynamic X-ray computed tomography (CT) enables time-resolved three-dimensional imaging of rapidly evolving processes, but its full potential is often constrained by the lack of real-time reconstruction and analysis capabilities. Modern high-speed detectors generate projection data at rates that far exceed the throughput of conventional three-dimensional reconstruction pipelines, posing significant challenges in data transmission, computation, and memory management. In this work, we present a live reconstruction and processing framework designed for ultra-fast dynamic CT experiments. The proposed framework integrates a high-speed data transfer module implemented using Remote Direct Memory Access, a multi-stage CUDA stream design, and a GPU-accelerated analytical reconstruction algorithm to bridge the gap between high-rate data acquisition and real-time volumetric imaging. By partitioning large volumes into subvolumes and carefully orchestrating host-to-device transfers, kernel execution, and device-to-host transfers, the framework achieves sustained high throughput while avoiding GPU memory exhaustion. Benchmark results demonstrate that the proposed framework can continuously process large volumes at projection consumption rates compatible with modern X-ray detectors exceeding 10 GB/s. These results indicate that the proposed system provides a practical and scalable solution for live three-dimensional reconstruction in ultra-fast dynamic CT experiments.