Semi-supervised low-dose SPECT restoration using sinogram inner-structure aware graph neural network.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Low-dose single photon emission computed tomography (SPECT) is crucial for reducing radiation risk.
• Current deep learning methods often overlook sinogram inner-structure and require extensive labeled data.
• Acquiring normal-dose SPECT data for supervised learning is challenging.

Purpose of the Study:

• To develop a semi-supervised framework for low-dose SPECT sinogram restoration.
• To exploit the inherent inner-structure of sinograms using graph neural networks.
• To leverage abundant unlabeled low-dose data to improve image quality.

Main Methods:

• A UNet-based framework incorporating sinogram-structure-based non-local neighbors graph neural network (SSN-GNN) and window-based K-nearest neighbors GNN (W-KNN-GNN).
• Utilized the mean teacher semi-supervised learning approach for training.
• Employed XCAT anthropomorphic digital phantoms for data generation.

Main Results:

• The proposed framework demonstrated superior performance over state-of-the-art methods in quantitative and qualitative evaluations.
• Ablation studies confirmed the effectiveness of each component, and robustness experiments showed resilience to varying noise levels.
• The semi-supervised approach effectively utilized unlabeled data to enhance restoration.

Conclusions:

• The developed framework effectively improves low-dose SPECT image quality by leveraging sinogram inner-structure and unlabeled data.
• This approach offers a valuable tool for radiation dose reduction in SPECT imaging without compromising image quality.
• The sinogram inner-structure-aware and semi-supervised strategy represents a significant advancement in low-dose SPECT reconstruction.