An Image Denoising Model using Deep Learning for Digital Breast Tomosynthesis Images

Breast cancer refers to a group of breast tumor subtypes with different genetic and cellular origins and clinical characteristics. Digital Breast Tomosynthesis (DBT) is a new technology that can help radiologists identify breast cancer more accurately. DBT method produces a series of low-dose full-field projection images with a lower acquisition angle, reconstruction noise and artifacts are expected. Thus, this study proposes a technique to denoise the DBT images using a deep learning model. The DBT images are obtained from the website, VICTRE trial, that host the clinical trial image for evaluating DBT. The neural network that is being studied will use a multiscale context aggregation network (MS-CAN) to learn to denoising the images. The deep learning method that is used is compared to the conventional method which is bilateral filtering. Statistical measures such as peak signal to noise ratio (PSNR), structural similarity index (SSIM), and Naturalness Image Quality Evaluator (NIQE) were used to evaluate the performance of the method.