Trustworthy Limited Data CT Reconstruction Using Progressive Artifact Image Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Limited data computed tomography (CT) aims to improve image quality with fewer projections, reducing radiation exposure and scan time.
• Deep Learning (DL) methods show promise but struggle with data distribution inconsistencies and lack of kernel awareness, limiting generalization and stability.
• Existing DL approaches often fail to adequately address artifacts arising from sparse or limited-angle CT data.

Purpose of the Study:

• To propose a novel unrolling framework, Progressive Artifact Image Learning (PAIL), for high-quality limited data CT reconstruction.
• To enhance the generalization and stability of DL-based CT reconstruction methods.
• To overcome limitations of current DL techniques, including distribution inconsistency and lack of kernel awareness.

Main Methods:

• Developed the Progressive Artifact Image Learning (PAIL) framework, an unrolling approach for limited data CT reconstruction.
• Incorporated three key modules: Residual Domain Module (RDM) for artifact suppression in residual images, Image Domain Module (IDM) for artifact reduction in the image space, and Wavelet Domain Module (WDM) for enhanced stability and kernel awareness.
• Integrated wavelet-based compressed sensing within the WDM to make the network kernel-aware and prevent hallucinations.

Main Results:

• The PAIL framework demonstrated superior performance in limited data CT reconstruction tasks across simulated, clinical cardiac, and sheep lung datasets.
• PAIL effectively suppressed observable and unobservable artifacts, improving image quality compared to state-of-the-art methods.
• The proposed method showed promising generalization capabilities and enhanced stability, outperforming existing techniques.

Conclusions:

• The Progressive Artifact Image Learning (PAIL) framework offers a robust solution for limited data CT reconstruction, significantly improving image quality and stability.
• PAIL effectively addresses the challenges of distribution inconsistency and lack of kernel awareness in DL-based CT reconstruction.
• The method's strong performance across diverse datasets highlights its potential for clinical applications requiring reduced radiation exposure or faster scanning.