上海科技大学知识管理系统

Purpose: The metal artifacts in the CT image detrimentally affect clinical diagnosis. The dual-domain (sinogram domain and image domain) networks have shown great potentials in reducing metal artifacts. The goal of this study is to develop a dual-domain deep learning network to reduce undesirable metal artifacts in CT.

Methods: The proposed deep learning method leverage information from dual domains and is embedded in an iterative framework. In each iteration, the metal-trace corrupted sinogram is first passed through the sinogram domain network to receive corrected sinogram output, then the output sinogram is FBP reconstructed to receive its corresponding image, and finally this image is enhanced by another network in image domain. The iterative number is 10. The sinogram domain network uses a prior sinogram, which is generated in advance by a prior network. In addition, the output sinogram from the sinogram network is further optimized by a sinogram domain fidelity module (SDFM). The proposed network is carefully trained using the public clinical dataset (DeepLesion) and its performance is evaluated both qualitatively and quantitatively.

Results: Compared to the state-of-the-art methods, the proposed method achieves improved metal artifact correction. The PSNR and SSIM is improved by 10.7% and 4%, respectively. The visual comparison demonstrates that our method can recover the image structure even with sever metal artifacts. In addition, the proposed model preserves small structures and avoids secondary artifacts.

Conclusion: Preliminary results suggest that the proposed network can achieve significantly better performance in metal artifact correction in CT.