Lightweight Network Enhancing High-Resolution Feature Representation for Efficient Low Dose CT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Low-dose computed tomography (CT) is essential for reducing radiation exposure but suffers from significant image noise, compromising diagnostic accuracy.
• Transformer-based models show promise for CT denoising but often exhibit high computational complexity, limiting their clinical applicability.

Purpose of the Study:

• To develop a lightweight and computationally efficient deep learning network, AMFA-Net, for enhancing image quality in low-dose CT.
• To improve high-resolution feature representation and robust denoised image reconstruction.

Main Methods:

• Proposed AMFA-Net, an adaptive multi-order feature aggregation network with a lightweight architecture.
• Introduced an agent-based self-attention cross-shaped window transformer block for efficient global context capture in high-resolution feature maps.
• Employed multi-order gated aggregation to adaptively capture expressive interactions and preserve structural information.

Main Results:

• AMFA-Net demonstrated superior denoising performance compared to state-of-the-art methods on two public datasets, achieving high image quality at 25% and 10% of full-dose CT.
• The proposed method achieved significant noise reduction while preserving critical structural information.
• The network operates with low computational cost, indicating potential for real-time applications.

Conclusions:

• AMFA-Net offers an effective and efficient solution for low-dose CT denoising, enhancing image quality and diagnostic precision.
• The lightweight architecture and adaptive feature aggregation enable robust denoised image reconstruction with reduced computational burden.
• This approach holds significant promise for advancing real-time medical imaging applications.