Unsupervised Learnable Sinogram Inpainting Network (SIN) for Limited Angle CT reconstruction

TL;DR

This paper introduces a neural network-based sinogram inpainting method for limited-angle CT reconstruction, combining physical model integration and unsupervised learning to improve image quality with limited data.

Contribution

It proposes a novel unsupervised learnable sinogram inpainting network (SIN) that integrates physical models and enables self-training for limited-angle CT reconstruction.

Findings

Outperforms SART-TV in PSNR and SSIM metrics

Enables unsupervised domain adaptation with limited data

Provides significant visual improvements in reconstructed images

Abstract

In this paper, we propose a sinogram inpainting network (SIN) to solve limited-angle CT reconstruction problem, which is a very challenging ill-posed issue and of great interest for several clinical applications. A common approach to the problem is an iterative reconstruction algorithm with regularization term, which can suppress artifacts and improve image quality, but requires high computational cost. The starting point of this paper is the proof of inpainting function for limited-angle sinogram is continuous, which can be approached by neural networks in a data-driven method, granted by the universal approximation theorem. Based on this, we propose SIN as the fitting function -- a convolutional neural network trained to generate missing sinogram data conditioned on scanned data. Besides CNN module, we design two differentiable and rapid modules, Radon and Inverse Radon Transformer network, to encapsulate the physical model in the training procedure. They enable new joint loss functions to optimize both sinogram and reconstructed image in sync, which improved the image quality significantly. To tackle the labeled data bottleneck in clinical research, we form a sinogram-image-sinogram closed loop, and the difference between sinograms can be used as training loss. In this way, the proposed network can be self-trained, with only limited-angle data for unsupervised domain adaptation. We demonstrate the performance of our proposed network on parallel beam X-ray CT in lung CT datasets from Data Science Bowl 2017 and the ability of unsupervised transfer learning in Zubal's phantom. The proposed method performs better than state-of-art method SART-TV in PSNR and SSIM metrics, with noticeable visual improvements in reconstructions.