Digital chest tomosynthesis: The main steps to a computer assisted lung diagnostic system

Abstract

In this paper, we present the main parts of a complete lung diagnostic system using digital tomosynthesis, and the first results obtained analyzing real tomosynthesis (DTS) images. In a DTS system first coronal image slices are reconstructed from projections using iterative and MITS reconstruction algorithms. Nodule detection is based on 2D image processing on the separated slice images, and a joint further analysis of the 2D results. We propose efficient, domain-specific filters for the enhancement and classification of bright, rounded structures. Also we develop a vessel enhancing algorithm based on strain energy filters. Vessel enhancement is required because most of the false positive findings come from nodule-like vessel shadows in the images. To reduce false positive findings SVM-based classifiers are applied, where features obtained from the vessel enhancement module are used as inputs. The system was evaluated on the first DTS scans, obtained from our experimental DTS system. The database contains ~2000 nodule candidates. 97% of nodules could be detected, while producing on average 31 false positives per scan.