A dual-interactive fusion network for low-dose CT image denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose computed tomography (LDCT) is crucial for reducing radiation exposure in clinical imaging.
• However, LDCT images suffer from quantum noise and artifacts, degrading image quality and diagnostic accuracy.
• Existing denoising methods like model-based iterative denoising (MBIR) are computationally intensive, while deep learning approaches lack generalizability.

Purpose of the Study:

• To develop an advanced deep learning framework, DIFNet, for effective low-dose computed tomography (LDCT) image denoising.
• To address the limitations of existing methods by improving noise reduction, artifact mitigation, and generalization capabilities.
• To enhance diagnostic accuracy and reliability in LDCT imaging.

Main Methods:

• Proposed the Dual-Interactive Fusion Network (DIFNet) framework for LDCT image denoising.
• Integrated the Dual-Phase Denoising Architecture (DPDA) and Context-Aware Training Strategy (CATS).
• Utilized a combined dual-phase loss function to optimize denoising performance.

Main Results:

• DIFNet demonstrated significant noise and artifact reduction on in-house LDCT datasets from Philips scanners.
• The framework preserved essential anatomical details, outperforming established methods like REDCNN, EDCNN, DDPM, and CTformer.
• Validated robust and competitive performance on the public Mayo-2016 benchmark (Siemens scanner) and through ablation studies.

Conclusions:

• DIFNet shows significant potential for clinical application in LDCT denoising, improving diagnostic reliability.
• The framework offers a balance of performance, robustness, and computational efficiency.
• DIFNet advances the field of LDCT denoising, contributing to better patient care.