Linear diffusion noise boosted deep image prior for unsupervised sparse-view CT reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning significantly improves sparse-view computed tomography (CT) reconstruction.
• Supervised methods require extensive paired datasets and retraining for new imaging conditions, limiting generalizability.

Purpose of the Study:

• To develop an unsupervised deep learning approach for sparse-view CT reconstruction.
• To enhance generalizability and adaptability without reliance on paired training data.

Main Methods:

• Utilized a deep image prior framework with multi-level linear diffusion noise to prevent overfitting.
• Incorporated non-local self-similarity within a self-attention network for pattern recognition.
• Employed gradient backpropagation between image and projection data spaces to optimize network weights.

Main Results:

• Demonstrated effective zero-shot adaptability across diverse projection views in simulated and clinical cases.
• Successfully eliminated noise and streak artifacts.
• Significantly restored intricate image details, showcasing robustness and flexibility.

Conclusions:

• The proposed unsupervised method overcomes limitations of supervised deep learning in sparse-view CT.
• Offers improved generalizability and adaptability for CT reconstruction without extensive paired training data.