Fast and efficient fully 3D PET image reconstruction using sparse system matrix factorization with GPU acceleration

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Iterative image reconstruction is crucial for Positron Emission Tomography (PET) imaging.
• Accurate system matrices are vital for image quality but computationally expensive and require large storage.
• Existing methods face challenges with high-resolution imaging due to computational demands.

Purpose of the Study:

• To develop an efficient method for PET image reconstruction.
• To reduce the computational cost and storage requirements of accurate system matrices.
• To enhance the performance of high-resolution iterative image reconstruction.

Main Methods:

• Sparse matrix factorization of the system matrix into three smaller, sparse matrices.
• GPU acceleration for efficient forward and backward projection implementation.
• Estimation of sinogram and image blurring matrices from point-source measurements.

Main Results:

• The proposed method dramatically reduces computation cost for high-resolution PET image reconstruction.
• The factored system matrix significantly lowers storage and computation demands.
• Achieved better performance compared to existing factorization techniques.

Conclusions:

• Sparse matrix factorization combined with GPU acceleration offers an efficient solution for PET image reconstruction.
• This approach overcomes the limitations of high computation cost and storage requirements.
• Enables improved performance in high-resolution PET imaging.