A Hybrid Alternating Direction Method of Multipliers Based on Structural Guided Filtering for the Reconstruction of Robust Low Dose CT

Abstract

Reduced radiation exposure in the process of computed tomography (CT) needs to be kept down in order to protect the health of patients. Automated and deep learning methods offer rapid, efficient reconstructions of higher-quality low-dose CT images. But these control images increasingly show that, due to undesired artifacts, quantum noise and streak interference in large quantities are then inflicted upon their visual content as well. Filtered Back-Projection (FBP), a classical reconstruction method in low-dose diagnostics, simply does not work very well. Pure deep learning algorithms, however, always result in pictures with poor visibility that destroy complex structural features. Here, we provide a systematic solution that combines physical data regularity and deep learning priors. The Alternating Direction Method of Multipliers (ADMM) is used to define the reconstruction problem. Then, as a regularization term, the trained Denoising Convolutional Neural Network (CDNN) can further enhance the images (remember our main goal is reducing noise). At this point, an adaptive Guided Filtering step is further included that, based on the structures learned in the initial reconstruction, preserves edges and low-frequency information in order to ensure that the deep denoising does not completely lose texture. Experimental results on a clinical dataset show that the proposed approach attains a PSNR (average peak signal-to-noise ratio) of 31.04 dB, with rich texture, which is used for diagnosis (quality SSIM). In conclusion, the net reduction in noise combined with improved chances for a diagnosis will be a boon for hospital personnel compared with existing approaches. These results demonstrate the proposed hybrid method's ability to successfully eliminate noise while maintaining crucial diagnostic data.