Digital tomosynthesis mammography (DTM) is a new modality that holds the promise of improving mammographic sensitivity of breast cancer detection and diagnosis, especially for dense breasts. The main goals of the proposed research are (1) to develop a computer-aided detection (CADd) system for breast masses in DTM, (2) to develop a computer-aided diagnosis (CADx) system for classification of malignant and benign masses in DTM, and (3) to evaluate the effects of CAD (either CADd or CADx) on radiologists' interpretation of DTMs. Previous CAD systems are developed for regular projection mammograms (PMs).The proposed CAD system makes use of the 3-dimensional (3D) information in DTM to improve mass detection and characterization. The innovations in the proposed project include: (1) development of new computer-vision techniques to exploit the 3D volumetric information in DTMs, (2) evaluation of the dependence of CAD performance on reconstruction algorithms, and (3) comparison of computerized mass detection and characterization in DTMs, projection view mammograms (PVs) (the non-reconstructed mammograms taken at multiple angles during tomosynthesis imaging), and PMs. We hypothesize that detection and characterization of masses on DTMs will be more accurate than corresponding tasks on regular PMs, and that the CAD systems can improve radiologists' accuracy. 3D breast phantoms with test objects will be designed and imaged with a prototype DTM system. The dependence of DTM image quality on reconstruction algorithms and their parameters, and on image acquisition techniques will be studied. The appropriate reconstruction techniques will be selected based on phantom and patient studies. A database of DTMs and corresponding PMs with malignant and benign masses and a set of normal cases will be collected with patient informed consent. CAD systems for detection and classification of masses will be developed. Two approaches will be compared: one uses the reconstructed DTM slices and the other uses the PVs as input to the CAD systems. For the DTMs, new techniques for 3D preprocessing, image segmentation, feature extraction, and feature classification will be designed. For the PVs, our previous techniques developed for regular PMs will be adapted to these low- dose images, and information fusion methods using techniques such as neural networks or support vector machines will be developed to merge the multiple-PV information. To test our hypotheses, we will compare the CAD system performances from these two approaches and that from the corresponding regular PMs, and conduct observer ROC studies to evaluate effects of the CAD systems on radiologists' performance. CAD will be an important tool that can help accelerate the implementation of DTM in clinical practice. DTM with CAD is expected to help fully utilize the potential of this new modality to improve breast cancer detection. ? ? ?
Showing the most recent 10 out of 11 publications
