Low-dose three-phase brain perfusion imaging and AI-based parameter map generation:Study

Now, team from the First Affiliated Hospital of Jinan University and Southern Medical University, have developed an innovative CTP scanning protocol and a deep learning model that can generate the vital blood flow maps needed to assess stroke patients. This work has proofed the proposed low radiation dose imaging program can slash radiation exposure by over 80% compared to current methods. This innovation promised to make stroke diagnosis safer and more accessible, particularly for vulnerable patients.

Addressing Limitations in Conventional CTP

Despite its clinical value, traditional CTP is associated with significant drawbacks. These include:

1. High Radiation Dose: Conventional protocols can reach cumulative doses around 5260 mGy·cm, notably higher than CTA (~3222 mGy·cm).
2. Motion Sensitivity: Repeated scans across timepoints make the technique vulnerable to patient motion, requiring sophisticated registration algorithms.
3. Workflow Complexity: The image processing burden and risk of failure hinder routine use in clinical settings.

Previous strategies for reducing radiation via temporal subsampling risk omitting critical arterial enhancement peaks, underestimating hemodynamic parameters. While multiphase CTA (mCTA) has shown promise in capturing arterial and venous phases, it requires large contrast volumes (~80 mL), posing risks for patients with impaired renal function, and lacks quantitative perfusion data.

A Three-Phase CTP with Deep Learning Enhancement

Inspired by the temporal structure of mCTA, the team introduced a three-phase CTP protocol that drastically reduces temporal sampling while preserving essential perfusion information. A generative adversarial network (GAN)-based model was developed to directly synthesize perfusion parameter maps from only three timepoints.

In internal validation datasets, the model-produced maps showed high structural and perceptual fidelity compared to ground truth, demonstrating its capability to reconstruct key perfusion features. Further experiments explored how variations in the selected three-phase combinations affected performance. Even with ±2-second deviations from the ideal timepoints, the model maintained high predictive accuracy, although performance dropped with deviations beyond 4 seconds. These findings support both the practical feasibility of the protocol and the robustness of the model.

Reference:

Cuidie Zeng, Xiaoling Wu, Fusheng Ouyang, Baoliang Guo, Xiao Zhang, Jianghua Ma, Dong Zeng, Bin Zhang. Perfusion Parameter Map Generation from 3 Phases of Computed Tomography Perfusion in Stroke Using Generative Adversarial Networks. Research. 2025;8:0689.DOI:10.34133/research.0689.