Effect of denoising on the quality of reconstructed images in digital breast tomosynthesis

Abstract

Individual projection images in Digital Breast Tomosynthesis (DBT) must be acquired with low levels of radiation, which significantly increases image noise. This work investigates the influence of a denoising algorithm and the Anscombe transformation on the reduction of quantum noise in DBT images. The Anscombe transformation is a variance-stabilizing transformation that converts the signal-dependent quantum noise to an approximately signalindependent Gaussian additive noise. Thus, this transformation allows for the use of conventional denoising algorithms, designed for additive Gaussian noise, on the reduction of quantum noise, by working on the image in the Anscombe domain. In this work, denoising was performed by an adaptive Wiener filter, previously developed for 2D mammography, which was applied to a set of synthetic DBT images generated using a 3D anthropomorphic software breast phantom. Ideal images without noise were also generated in order to provide a ground-truth reference. Denoising was applied separately to DBT projections and to the reconstructed slices. The relative improvement in image quality was assessed using objective image quality metrics, such as peak signal-to-noise ratio (PSNR) and mean structural similarity index (SSIM). Results suggest that denoising works better for tomosynthesis when using the Anscombe transformation and when denoising was applied to each projection image before reconstruction; in this case, an average increase of 9.1 dB in PSNR and 58.3% in SSIM measurements was observed. No significant improvement was observed by using the Anscombe transformation when denoising was applied to reconstructed images, suggesting that the reconstruction algorithm modifies the noise properties of the DBT images.