Dynamic PET Image Reconstruction Using Kalman Inspired Network

Positron Emission Tomography (PET) facilitates the visualization of the distribution of radioactive isotope-labeled compounds within biological organisms. Existing reconstruction algorithms, including iterative and deep learning methods that depend on the system matrix, often suffer from inaccuracies in the system matrix, resulting in artifacts or blurring in the reconstructed images. Inspired by Kalman Filtering, we propose a data-driven deep learning framework for PET image reconstruction, named the Kalman Inspired Network (KIN). The KIN framework divides the reconstruction problem into two phases: state prediction and state update, and consists of three core components: Prediction Net, Projection Net, and Kalman Gain Net. By adopting a data-driven approach, KIN circumvents the reliance on the system matrix, thereby overcoming the limitations imposed by noise prior knowledge and the inversion of large-dimensional matrices typically required in traditional Kalman Filtering algorithms. We evaluated the proposed KIN network using simulated phantom and experimental rat datasets, benchmarking it against traditional algorithms as well as other deep learning-based methods. The results demonstrate that KIN enhances the quality of dynamic PET scans, both in terms of image quality and quantitative indices, underscoring its potential for dynamic imaging applications that require extremely short frames and are typically challenged by high noise levels.