Novel image reconstruction approaches to improving image quality of digital breast tomosynthesis

Abstract

Digital breast tomosynthesis (DBT) is an emerging breast cancer screening and diagnostic tool which has the advantages over digital mammography. DBT can produce quasi three-dimensional images from the projections mitigating an inherent limitation of tissue overlap in mammography. Since the detection of lesions such as micro-calcifications and mass in breasts is the purpose of using DBT, it is desirable to develop a technique producing higher detectability of lesions. DBT images, however, often suffer from high-density object artifacts due to its incompleteness of data. Ripple artifacts may appear in the out-of-focus planes due to limited angle scan. The ripple artifacts occur more seriously with high-density objects. Moreover, undershoot artifacts that show up as dark fringes near the high-density object border are quite overwhelming in the reconstructed images by filtered back-projection (FBP) algorithm. Because high-density object artifacts may obscure significant anatomical structures in DBT not only for diagnosis and screening but also for DBT-related applications such as DBT-guided breast biopsy, an algorithm mitigating high-density objects artifacts would be required. In this thesis, we developed a pre-processing technique for DBT that results in a higher detectability and less typical breast border enhancement artifacts than a conventional image reconstruction. A histogram modification technique was developed in the projection data domain. Histogram of raw projection data was first divided into two parts: One for the breast projection data and the other for background. Background pixel values were set to a single value that represents the boundary between breast and background. After that, both histogram parts were shifted by an appropriate amount of offset and the histogram-modified projection data was log-transformed. FBP algorithm was used for image reconstruction of DBT. To evaluate performance of the proposed method, we computed the detectability index for the reconstructed images from clinically acquired data. Breast border enhancement artifacts were greatly suppressed and the detectability of calcifications and masses was increased by use of the proposed method. In addition, we address high-density object artifacts problem by use of back-projection filtration (BPF) reconstruction algorithm. Data derivatives were back-projected with different weights to reduce ripple artifacts by use of a voting strategy. We generated another differentiated back-projection volume where edges of high-density objects are replaced by the background to reduce the undershoot artifacts. After applying the Hilbert transform, we blended the two images. For evaluation, we calculated artifacts volume fraction (AVF). We set the volume of interests that are contaminated by the artifacts, and segmented the artifacts volume. We defined AVF as ratio of artifacts volume to total volume of VOI. CIRS breast phantom and a lab-made pork phantom mimicking DBT-guided breast biopsy were scanned. We compared three image reconstruction methods: conventional FBP algorithm, FBP utilizing a voting strategy, and the proposed method. Both ripple artifacts and undershoot artifacts were greatly suppressed by the proposed method. The proposed method resulted in AVF values about 75 % less than those in FBP reconstructions. We have introduced novel image reconstruction approaches to improving image quality of DBT image. Our algorithms are believed to play important roles in many applications of digital tomosynthesis although we have focused on DBT only in this thesis.