Anatomically adaptive feature-wise linear modulation for deep learning-based low-dose CT denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces heterogeneous noise.
• Existing deep learning methods often assume uniform noise, leading to suboptimal performance in diverse anatomical regions.
• There is a need for noise reduction techniques that account for region-dependent noise characteristics in LDCT.

Purpose of the Study:

• To design and evaluate a physics-based framework for anatomically adaptive noise removal in LDCT.
• To explicitly model and address region-dependent noise characteristics for improved LDCT image quality.
• To develop a deep learning model that balances noise removal with the preservation of anatomical details.

Main Methods:

• Proposed an anatomically adaptive noise reduction framework utilizing a U-Net architecture.
• Integrated Global Feature-wise Linear Modulation (GFiLM) for global noise removal and Local FiLM (LFiLM) for tissue-specific noise reduction.
• Evaluated the Anatomically Adaptive Feature-wise Linear Modulation (AA-FiLM) model on AAPM-Mayo Clinic and TCIA LDCT datasets.

Main Results:

• The AA-FiLM model outperformed existing methods in LDCT noise removal.
• Achieved consistent noise reduction across lung (37.14%), soft tissue (48.75%), and bone (35.18%) regions.
• Demonstrated significant improvements in noise removal, structural detail preservation, and reduced distortions, with good generalization capabilities.

Conclusions:

• The developed framework enables anatomically adapted noise removal by considering physical noise properties and tissue types.
• The dual-modulation approach effectively addresses both general and tissue-specific noise in LDCT.
• The AA-FiLM model offers a robust solution for LDCT noise removal, balancing image quality and diagnostic detail.