Computer-aided detection for digital breast tomosynthesis

Digital Breast Tomosynthesis (DBT) is a new three-dimensional limited-angle tomography breast imaging technique that will substantially overcome the superimposition problem for lesion detection in digital mammography. This work focuses on developing an automated detection scheme for the very recent images obtained with DBT. Since DBT is an emerging modality, clinical data sets are still rare, and few works have been published in the domain of computer-aided detection for these data sets. The three dimensional nature of the data offers a number of advantages. At the same time, issues like a reduced patient dose per scan, the limited-angle acquisition geometry and the resulting reconstruction artifacts, as well as the problem of an adapted reconstruction technique need to be addressed. We develop an algorithm for reconstruction-independent computer-aided detection on DBT data. Working directly on the tomographic projection images, we make use of fuzzy set theory to preserve the ambiguity present in the images until the information from the entire set of projected views can be jointly considered. A framework is proposed that allows for detection of different kinds of radiological findings using the same general framework. We present a method for fuzzy contour extraction for mass lesions. The originality of our method lies in the use of a set of dedicated hybrid active contour models. Each model is constructed using a priori knowledge about a class of objects of interest. For each radiological finding, we apply the different active contour models resulting in a set of fuzzy contours for each object of interest. A feature vector is extracted for each fuzzy contour. Using the extension principle we establish a set of fuzzy attributes based on this feature vector. We propose an aggregation strategy for combining the information corresponding to a given three-dimensional object over the entire set of projected views. To this end we introduce a partial defuzzification operation that provides a pixel-based representation of the fuzzy information extracted for each object. This operation, combined with a back-projection/re-projection step, provides the links between the particles in the projection images, that correspond to the same three-dimensional particle. Cumulated fuzzy attributes and a confidence degree for each particle are computed. A fuzzy decision tree working on fuzzy data is then applied to obtain a final decision. The presented approach has been validated on a number of simulated images as well as on clinical images. We show that the algorithm is capable of distinguishing between different radiological signs. The availability of a larger clinical database will allow a better quantitative evaluation of the algorithm performance in a near future.