Deep-unfolding-network-based non-blind deblurring for fast-rotating wide-angle digital breast tomosynthesis

Abstract

Digital breast tomosynthesis is a pseudo 3d imaging technique that uses only limited scan angle and has been recently used for breast cancer screening and diagnosis. There are many advantages of DBT. It can compensate for the tissue overlap issues, which is a weakness of X-ray mammography, and has a lower breast cancer false-positive rate and false-negative rate than that of X-ray mammography. It is very important to determine the angular range of the DBT scan. Wide-angle DBT (about 50°) has a higher contrast and better separability of overlapping tissues than narrow-angle DBT (about 15°). But, there is a significant disadvantage in that the scan time. Because a long scan time not only increases the discomfort caused by breast compression paddles for patients but also causes the patient motion blur. So, a fast-rotating X-ray tube is desirable to reduce the wide-angle DBT scan time. However, it can cause another artifact which is source motion blur. It can severely degrade the image quality, making for the reader difficulty to identify the lesion. In this work, we address this problem and propose a deblurring methodology. First, we modeled an analytic simulation system for a high-speed wide-angle DBT and estimated the effect of source motion blur and blur kernel. Next, to correct the effect of source motion blur, we implemented a non-blind-based deep-unfolding network that performs deblurring tasks using a blur kernel. Our proposed methodology show it can correct the effect of source motion blur. We highly value the clinical impact of the proposed method since it can alleviate the discomfort caused by prolonged breast compression in wide-angle DBT scans.