A Two-Branch Neural Network for Short-Axis PET Image Quality Enhancement

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Axial field of view (FOV) is critical for PET image quality.
• Hardware limitations in conventional PET scanners result in low-quality PET (LQ-PET) images.
• Existing deep learning methods struggle with long-range contextual information extraction.

Purpose of the Study:

• To introduce a novel deep learning network, SW-GCN, for enhancing LQ-PET images.
• To overcome hardware FOV restrictions and improve PET image quality.
• To achieve better clinical diagnoses through superior PET imaging.

Main Methods:

• Proposed a two-branch network architecture: SW-GCN, integrating swin transformer units and graph convolution.
• Spatial adaptive branch uses windowed self-attention and shifted operations to manage computational cost.
• Channel adaptive branch employs Watts Strogatz topology for efficient feature map connection and reduced redundancy.
• Ensemble learning combines features from both branches for enhanced PET image reconstruction.

Main Results:

• SW-GCN demonstrated superior performance over state-of-the-art methods in quantitative and qualitative evaluations.
• Extensive experiments were conducted on 386 patients across three single-bed position scans.
• The model effectively extracts and fuses long-range contextual information, addressing limitations of previous approaches.

Conclusions:

• The proposed SW-GCN effectively enhances low-quality PET images by improving the extraction and fusion of contextual information.
• SW-GCN offers a promising solution for overcoming hardware limitations in PET imaging.
• This advanced deep learning approach has the potential to significantly improve clinical diagnoses through enhanced PET image quality.