Neural architecture search with Deep Radon Prior for sparse-view CT image reconstruction

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Sparse-view computed tomography (CT) reduces radiation exposure but introduces artifacts due to data scarcity.
• Existing deep learning (DL) methods for CT reconstruction often require extensive paired datasets and lack interpretability.

Purpose of the Study:

• To develop a novel unsupervised deep learning method for CT reconstruction.
• To leverage Deep Radon Prior (DRP) and Neural Architecture Search (NAS) for improved image fidelity.

Main Methods:

• Proposed NAS-DRP, an unsupervised DL method using reinforcement learning-based NAS for architectural optimization.
• Integrated data inconsistency in the Radon domain and insights from Deep Image Prior (DIP).
• Employed Recurrent Neural Networks (RNNs) to constrain optimization for sparse-view CT.

Main Results:

• NAS-DRP demonstrated significant performance improvements over traditional and other DL methods.
• Achieved superior objective metrics (PSNR, SSIM, LPIPS) and subjective visual quality.
• Effectively minimized artifacts and enhanced image detail and accuracy.

Conclusions:

• NAS-DRP advances CT image reconstruction by integrating NAS with DL and Radon domain adaptations.
• Addresses sparse-view CT challenges, reducing data acquisition costs and complexity.
• Shows substantial potential for widespread application in medical imaging.