Breast mass detection in 3D reconstructed volume of digital breast tomosynthesis

Abstract

Digital breast tomosynthesis (DBT) is a new three-dimensional (3D) tomographic imaging modality for early detection of breast cancer. DBT is designed to alleviate the tissue overlap problem which inherent in conventional two-dimensional (2D) mammography by reconstructing a 3D breast volume from a series of projection view images obtained over limited angular range. To help radiologists in clinical realm, researches for developing a computer-aided detection (CAD) system have been reported. Typically, CAD system for detecting masses in 3D reconstructed volume largely consists of the following two processes: volume of interest (VOI) detection and false positive (FP) reduction using feature analysis. In order to develop the CAD system for DBT, limitation of 3D reconstructed volume should be considered. The 3D reconstructed volume only provides the quasi-3D structure information with limited resolution along depth direction due to the insufficient sampling in depth direction. The limited angular range and limited number of projection views also causes the problem of blur in our-of-focus plane. These limitations could seriously hamper conventional 3D image analysis technique for detecting masses. In this thesis, we proposed a novel mass detection approach for detecting masses in on 3D reconstructed volume to overcome above limitations. Firstly, to overcome the limited resolution along depth direction, we detect regions of interest (ROIs) on each reconstructed slices and separately utilizes depth directional information to combine the ROIs effectively. Furthermore, we measure the blurriness of each slice and select in-focus slices to extract features for resolving the degradation of performance caused from the blur in out-of-focus plane. Finally, features are extracted from selected in-focus slices and analyzed by support vector machines (SVMs) classifier to reduce FPs. Comparative experiments have been conducted on the clinical data set provided by a private hospita...