Real-World Low-Dose CT Image Denoising by Patch Similarity Purification

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Reducing radiation dose in CT scanning is crucial for patient health.
• Low-dose CT (LDCT) imaging reduces radiation exposure but introduces significant noise, impacting diagnostic accuracy.
• Existing LDCT denoising methods struggle with synthetic noise or spatial misalignment in real-world data pairs.

Purpose of the Study:

• To develop a robust method for creating high-quality training datasets for LDCT image denoising using real-world data.
• To address the limitations of synthetic noise and spatial misalignment in current LDCT denoising training approaches.
• To improve the accuracy and reliability of LDCT image denoising networks.

Main Methods:

• Introduced a Patch Similarity Purification (PSP) strategy for LDCT image denoising.
• PSP involves binarizing image patches from LDCT and NDCT pairs and calculating mask similarity.
• Selected training samples based on a mask similarity ratio threshold to ensure negligible misalignment.

Main Results:

• The PSP strategy successfully constructed large-scale datasets (Rabbit and Patient) of real-world LDCT/NDCT patch pairs with minimal misalignment.
• Experiments demonstrated the effectiveness of PSP in purifying training data.
• Networks trained on PSP-purified datasets showed improved LDCT image denoising performance.

Conclusions:

• The Patch Similarity Purification strategy is a valuable tool for creating high-quality training data for LDCT denoising.
• Utilizing real-world, aligned data significantly enhances the performance of LDCT denoising networks.
• This approach offers a practical solution for improving diagnostic accuracy in low-dose CT imaging.