Deep Fusion Network Based Sparse View CT Reconstructions for Clinical Diagnostic Scanners

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Sparse view CT scans offer reduced radiation exposure and faster scanning.
• Limited projections in sparse view CT can lead to ill-posed reconstruction problems and image artifacts.

Purpose of the Study:

• To introduce a novel model-based deep fusion network (DFN) for high-quality sparse view CT reconstruction.
• To improve image quality and reconstruction speed in clinical CT diagnosis.

Main Methods:

• Developed a Deep Fusion Network (DFN) that fuses features from sinogram data and preliminary Filtered Back Projection (FBP) images.
• Utilized a custom loss function during training to optimize for pixel accuracy and tissue structure integrity.
• Trained and validated the DFN on a synthetic sparse view breast CT dataset from the American Association of Physicists in Medicine (AAPM).

Main Results:

• The DFN effectively extracts fused features, incorporating prior knowledge from FBP images to enhance reconstruction quality.
• The custom loss function guided the network to learn both pixel values and structural integrity.
• Qualitative and quantitative evaluations demonstrated significant improvements in balancing image quality and reconstruction speed.

Conclusions:

• The proposed DFN algorithm enables fast and high-quality CT reconstruction, overcoming limitations of sparse view acquisition.
• DFN offers a promising solution for clinical applications requiring efficient and accurate CT imaging.