Optimization of sparse-view CT reconstruction based on convolutional neural network.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Sparse-view CT reduces scan time and radiation dose but introduces streak artifacts due to undersampling.
• Deep learning methods show promise in mitigating these artifacts and enhancing sparse-view CT image quality.

Purpose of the Study:

• To improve sparse-view CT reconstruction by enhancing deep learning optimization capabilities.
• To increase the interpretability of deep learning methods for artifact removal.
• To boost the generalization of reconstruction models across various sparse views.

Main Methods:

• Developed SRII-Net, a U-Net-based network with a copy pathway and residual image output block.
• Created diverse network structures to analyze layer contributions to artifact removal and interpretability.
• Utilized multiple datasets of varying sampling views for training and generalization testing.

Main Results:

• SRII-Net significantly outperforms existing networks, improving PSNR and SSIM metrics with millisecond-level optimization.
• Analysis revealed the critical roles of shallow (detail) and deep (abstract) layers in artifact suppression.
• Training with mixed datasets demonstrated enhanced optimization for diverse sparse-view reconstructions.

Conclusions:

• The proposed SRII-Net effectively suppresses artifacts in sparse-view CT images, enhancing generalization.
• The study provides deeper understanding of deep learning artifact removal, applicable to other imaging modalities.