Self-Supervised Denoising With Noise Propagation Model: Improving Material Decomposition in Photon-Counting CT.

Area of Science:

• Medical Imaging
• Physics
• Computer Science

Background:

• Photon-counting computed tomography (PCCT) enables material identification but decomposition techniques amplify noise.
• Existing denoising methods address degraded images, not the fundamental photon detection noise.
• Noise in PCCT material decomposition limits diagnostic accuracy and image quality.

Purpose of the Study:

• To develop a novel method combining physics-based noise analysis and deep learning for noise control during PCCT material decomposition.
• To improve material accuracy and virtual monochromatic image quality in PCCT.
• To create a flexible and practical solution for clinical settings.

Main Methods:

• Developed a physics-based noise analysis model linking detector noise to material-specific decomposition noise patterns.
• Implemented a self-supervised deep learning training strategy using probability-based optimization for efficient learning with limited data.
• Created an adaptive image improvement system for diverse scanning conditions and patient anatomies.

Main Results:

• The proposed method effectively controls noise during material decomposition, preserving material accuracy.
• Generated cleaner virtual monochromatic images compared to traditional methods.
• Demonstrated robust performance with small training datasets and adaptability to various clinical scenarios.

Conclusions:

• This research integrates theoretical noise analysis with deep learning for improved PCCT imaging.
• The method offers a practical solution for enhancing PCCT material decomposition and image quality in clinical practice.
• The approach balances noise reduction with material accuracy, advancing diagnostic capabilities.