LU-Net: combining LSTM and U-Net for sinogram synthesis in sparse-view SPECT reconstruction

Area of Science:

• Medical Imaging
• Radiological Sciences
• Computational Imaging

Background:

• Reducing radiation dose in single-photon emission computed tomography (SPECT) is crucial for patient safety.
• Low-dose SPECT can result in sparse projection data, leading to artifacts in reconstructed images.
• Existing neural network methods for sparse-view SPECT often overlook the sequential nature of projection data.

Purpose of the Study:

• To develop a deep learning architecture that leverages the sequence features of projection data for improved sparse-view SPECT reconstruction.
• To synthesize full-view projection data from sparse-view data, thereby mitigating artifacts and enhancing image quality.
• To evaluate the performance of the proposed method against existing techniques.

Main Methods:

• A novel network architecture, LU-Net, combining Long Short-Term Memory (LSTM) and U-Net, was proposed.
• LU-Net learns the mapping from sparse-view to full-view projection data, utilizing sequential information from adjacent view angles.
• Projection data were simulated using SIMIND software at varying sparsity levels (1/2, 1/3, 1/4 views).
• Reconstruction was performed using the preconditioned alternating projection algorithm (PAPA) on synthesized data.

Main Results:

• The LU-Net model demonstrated significant improvements in both global and local image quality metrics.
• Performance was evaluated across three levels of view sparsity, showing consistent benefits.
• LU-Net outperformed standard U-Net and traditional iterative reconstruction methods (TV-PAPA).

Conclusions:

• The proposed LU-Net architecture effectively utilizes sequential projection data features for sparse-view SPECT.
• LU-Net offers a promising approach for high-quality SPECT imaging at reduced radiation doses.
• This method has the potential to enhance diagnostic accuracy in low-dose SPECT applications.