Development of a denoising convolutional neural network-based algorithm for metal artifact reduction in digital tomosynthesis

Abstract

This study aimed to evaluate a denoising convolutional neural network reconstruction (DnCNNR) algorithm for reducing metal objects on digital tomosynthesis when using projection data for arthroplasty. For metal artifact reduction (MAR), we implemented a DnCNNR algorithm based on a training network (i.e., mini-batch stochastic gradient-descent algorithm with momentum) to estimate residual reference (140 keV virtual monochromatic [VM]) and object (70 kV with metal artifacts) images using projection data, and subtracted the estimated residual images from the object images using hybrid and subjectively reconstructed images (back projection and maximum likelihood expectation maximization [MLEM]). This DnCNNR algorithm was compared with a dual-energy material decomposition reconstruction algorithm (DEMDRA), VM, MLEM, established and commonly used filtered back projection (FBP) methods, and simultaneous algebraic reconstruction technique-total variation (SART-TV) with MAR processing. MAR was then compared using artifact index (AI) and texture analyses. For images that were in-focus were evaluated using a prosthesis phantom. The derived images yielded better results that were not influenced by the metal type (e.g., the AI was almost equal to the best value for the DEMDRA). The DnCNNR algorithm also yielded the best performance with regard in the texture analysis. The proposed algorithm is particularly useful for not affected by tissue misclassification.