Noise Characteristics Modeled Unsupervised Network for Robust CT Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Deep learning (DL) shows promise for computed tomography (CT) image reconstruction.
• Current DL methods are typically evaluated on data from identical scan protocols, limiting real-world applicability.
• Protocol variations (e.g., region, kVp, mAs) introduce distribution shifts, impacting DL model performance.

Purpose of the Study:

• To analyze the robustness of DL-based CT reconstruction methods against protocol-specific distribution shifts.
• To develop a novel unsupervised CT reconstruction method that addresses these protocol-specific perturbations.
• To improve the reliability and generalizability of DL in CT imaging.

Main Methods:

• Investigated the sensitivity of DL reconstruction methods to variations in CT scan protocols (region, kVp, mAs).
• Introduced a Gaussian Mixture Model (GMM) based unsupervised CT reconstruction network (GMM-unNet).
• Utilized unpaired sinogram data and an expectation-maximization algorithm for network training and optimization.

Main Results:

• DL reconstruction methods exhibit sensitivity to protocol-specific perturbations, linked to noise distribution shifts.
• The proposed GMM-unNet method effectively handles noise distribution differences between training and testing datasets.
• Experiments on diverse datasets demonstrated superior qualitative and quantitative performance of GMM-unNet compared to existing methods.

Conclusions:

• DL-based CT reconstruction is vulnerable to scan protocol variations due to noise distribution changes.
• The GMM-unNet method offers a robust and effective solution for unsupervised low-dose CT reconstruction.
• This unsupervised approach enhances the reliability of DL methods in diverse clinical CT imaging scenarios.