M₃NAS: Multi-Scale and Multi-Level Memory-Efficient Neural Architecture Search for Low-Dose CT Denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces image noise, hindering diagnosis.
• Convolutional neural networks (CNNs) show promise for LDCT denoising, but current architectures may be suboptimal.
• Neural network architecture search (NAS) can optimize CNN performance.

Purpose of the Study:

• To apply NAS for the first time to LDCT denoising.
• To propose a multi-scale and multi-level memory-efficient NAS (M3NAS) for LDCT.
• To improve image quality and diagnostic accuracy in LDCT.

Main Methods:

• Developed M3NAS, a NAS method incorporating multi-scale feature fusion and hybrid cell/network-level architecture search.
• Designed M3NAS to be memory-efficient, reducing model parameters and increasing inference speed.
• Evaluated M3NAS on two distinct datasets using extensive experimental validation.

Main Results:

• M3NAS achieved superior denoising performance compared to state-of-the-art methods on LDCT images.
• The proposed method resulted in fewer model parameters and faster inference times.
• Validation confirmed the effectiveness of the multi-scale and multi-level architecture for LDCT denoising.

Conclusions:

• M3NAS represents a significant advancement in automated deep learning for LDCT image denoising.
• The developed architecture search strategy offers a more efficient and effective approach to improving LDCT image quality.
• This work paves the way for enhanced diagnostic capabilities through optimized AI models in medical imaging.