Multi-domain integrative Swin transformer network for sparse-view tomographic reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose X-ray imaging requires fewer projection views, often resulting in severe streak artifacts that degrade image quality.
• Reconstructing high-quality images from sparse-view data is a significant challenge in medical imaging.

Purpose of the Study:

• To develop and evaluate a novel deep learning network, the multi-domain integrative Swin transformer network (MIST-net), for improving image reconstruction from sparse-view X-ray data.
• To enhance image quality by reducing artifacts and preserving fine details and edges.

Main Methods:

• Developed MIST-net, a Swin transformer-based network incorporating multi-domain features (data, residual data, image, residual image).
• Integrated a data consistency module using residual data and residual image sub-networks to minimize interpolation and reconstruction errors.
• Incorporated a trainable edge enhancement filter for edge detection and preservation.
• Designed a high-quality reconstruction Swin transformer (Recformer) for capturing global image features.

Main Results:

• MIST-net demonstrated superior performance in reconstructing images from sparse-view data compared to existing methods.
• Experimental results on numerical and real cardiac clinical datasets (48 views) showed improved image quality.
• The proposed method effectively preserved small features and sharp edges, outperforming competitors.

Conclusions:

• MIST-net offers a robust solution for artifact reduction and image quality enhancement in low-dose, sparse-view X-ray imaging.
• The network's multi-domain integration and Swin transformer architecture contribute to its effectiveness in capturing both local and global image information.