Sparse-view statistical iterative head CT image reconstruction via joint regularization

It is a significant challenge to accurately reconstruct medical computed tomography CT images with important details and features. Reconstructed images always suffer from noise and artifact pollution because the acquired projection data may be insufficient or undersampled. In reality, some "isolated noise points" similar to impulse noise always exist in low-dose CT projection measurements. Statistical iterative reconstruction SIR methods have shown greater potential to significantly reduce quantum noise but still maintain the image quality of reconstructions than the conventional filtered back-projection FBP reconstruction algorithm. Although the typical total variation-based SIR algorithms can obtain reconstructed images of relatively good quality, noticeable patchy artifacts are still unavoidable. To address such problems as impulse-noise pollution and patchy-artifact pollution, this work, for the first time, proposes a joint regularization constrained SIR algorithm for sparse-view CT image reconstruction, named "SIR-JR" for simplicity. The new joint regularization consists of two components: total generalized variation, which could process images with many directional features and yield high-order smoothness, and the neighborhood median prior, which is a powerful filtering tool for impulse noise. Subsequently, a new alternating iterative algorithm is utilized to solve the objective function. Experiments on different head phantoms show that the obtained reconstruction images are of superior quality and that the presented method is feasible and effective.