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

BACKGROUND: Sparse-view computed tomography (CT) reduces radiation exposure but suffers from severe artifacts caused by insufficient sampling and data scarcity, which compromise image fidelity. Recent advancements in deep learning (DL)-based methods for inverse problems have shown promise for CT reconstruction but often require high-quality paired datasets and lack interpretability. PURPOSE: This paper aims to advance the field of CT reconstruction by introducing a novel unsupervised deep learning method. It builds on the foundation of Deep Radon Prior (DRP), which utilizes an untrained encoder-decoder network to extract implicit features from the Radon domain, and leverages Neural Architecture Search (NAS) to optimize network structures. METHODS: We propose a novel unsupervised deep learning method for image reconstruction, termed NAS-DRP. This method leverages reinforcement learning-based NAS to explore diverse architectural spaces and integrates reinforcement learning with data inconsistency in the Radon domain. Building on previous DRP research, NAS-DRP utilizes an untrained encoder-decoder network to extract implicit features from the Radon domain. It further incorporates insights from studies on Deep Image Prior (DIP) regarding the critical impact of upsampling layers on image quality restoration. The method employs NAS to search for the optimal network architecture for upsampling unit tasks, while using Recurrent Neural Networks (RNNs) to constrain the optimization process, ensuring task-specific improvements in sparse-view CT image reconstruction. RESULTS: Extensive experiments demonstrate that the NAS-DRP method achieves significant performance improvements in multiple CT image reconstruction tasks. The proposed method outperforms traditional reconstruction methods and other DL-based techniques in terms of both objective metrics (PSNR, SSIM, and LPIPS) and subjective visual quality. By automatically optimizing network structures, NAS-DRP effectively enhances the detail and accuracy of reconstructed images while minimizing artifacts. CONCLUSIONS: NAS-DRP represents a significant advancement in the field of CT image reconstruction. By integrating NAS with deep learning and leveraging Radon domain-specific adaptations, this method effectively addresses the inherent challenges of sparse-view CT imaging. Additionally, it reduces the cost and complexity of data acquisition, demonstrating substantial potential for broader application in medical imaging. The evaluation code will be available at https://github.com/fujintao1999/NAS-DRP/.