Michael Golatta1, Désirée Zeegers2, Konstantinos Filippatos3, Leah-Larissa Binder2, Alexander Scharf2, Geraldine Rauch4, Joachim Rom2, Florian Schütz2, Christof Sohn2, Joerg Heil2. 1. University Breast Unit, Im Neuenheimer Feld 440, 69120, Heidelberg, Germany. Michael.Golatta@med.uni-heidelberg.de. 2. University Breast Unit, Im Neuenheimer Feld 440, 69120, Heidelberg, Germany. 3. MeVis BreastCare Solutions GmbH & Co. KG A MeVis Medical Solutions Company, Universitaetsallee 29, 28359, Bremen, Germany. 4. Institute of Medical Biometry and Informatics, University of Heidelberg, 69120, Heidelberg, Germany.
Abstract
PURPOSE: This study aims at developing and evaluating a prototype of a lesion candidate detection algorithm for a 3D-US computer-aided diagnosis (CAD) system. METHODS: Additionally, to routine imaging, automated breast volume scans (ABVS) were performed on 63 patients. All ABVS exams were analyzed and annotated before the evaluation with different algorithm blob detectors characterized by different blob-radiuses, voxel-sizes and the quantiles of blob filter responses to find lesion candidates. Lesions found in candidates were compared to the prior annotations. RESULTS: All histologically proven lesions were detected with at least one algorithm. The algorithm with optimal sensitivity detected all cancers (sensitivity = 100 %) with a very low positive predictive value due to a high false-positive rate. CONCLUSIONS: ABVS is a new technology which can be analyzed by a CAD software. Using different algorithms, lesions can be detected with a very high and accurate sensitivity. Further research for feature extraction and lesion classification is needed aiming at reducing the false-positive hits.
PURPOSE: This study aims at developing and evaluating a prototype of a lesion candidate detection algorithm for a 3D-US computer-aided diagnosis (CAD) system. METHODS: Additionally, to routine imaging, automated breast volume scans (ABVS) were performed on 63 patients. All ABVS exams were analyzed and annotated before the evaluation with different algorithm blob detectors characterized by different blob-radiuses, voxel-sizes and the quantiles of blob filter responses to find lesion candidates. Lesions found in candidates were compared to the prior annotations. RESULTS: All histologically proven lesions were detected with at least one algorithm. The algorithm with optimal sensitivity detected all cancers (sensitivity = 100 %) with a very low positive predictive value due to a high false-positive rate. CONCLUSIONS: ABVS is a new technology which can be analyzed by a CAD software. Using different algorithms, lesions can be detected with a very high and accurate sensitivity. Further research for feature extraction and lesion classification is needed aiming at reducing the false-positive hits.