PURPOSE: We present a hybrid 2D-3D ultrasound (US) rigid registration method for navigated prostate biopsy that enables continuous localization of the biopsy trajectory during the exam. METHODS: Current clinical computer-assisted biopsy systems use either sensor-based or image-based approaches. We combine the advantages of both in order to obtain an accurate and real-time navigation based only on an approximate localization of the US probe. Starting with features extracted in both 2D and 3D images, our method introduces a variant of the iterative closest point (ICP) algorithm. Among other differences to ICP, a combination of both the euclidean distance of feature positions and the similarity distance of feature descriptors is used to find matches between 2D and 3D features. The evaluation of the method is twofold. First, an analysis of variance on input parameters is conducted to estimate the sensitivity of our method to their initialization. Second, for a selected set of their values, the target registration error (TRE) was calculated on 29,760 (resp. 4000) registrations in two different experiments. It was obtained using manually identified anatomical fiducials. RESULTS: For 160 US volumes, from 20 patients, recorded during routine biopsy procedures performed in two hospitals by six operators, the mean TRE was [Formula: see text] mm (resp. [Formula: see text] mm). CONCLUSION: This work allows envisioning further developments for prostate navigation and their clinical transfer.
PURPOSE: We present a hybrid 2D-3D ultrasound (US) rigid registration method for navigated prostate biopsy that enables continuous localization of the biopsy trajectory during the exam. METHODS: Current clinical computer-assisted biopsy systems use either sensor-based or image-based approaches. We combine the advantages of both in order to obtain an accurate and real-time navigation based only on an approximate localization of the US probe. Starting with features extracted in both 2D and 3D images, our method introduces a variant of the iterative closest point (ICP) algorithm. Among other differences to ICP, a combination of both the euclidean distance of feature positions and the similarity distance of feature descriptors is used to find matches between 2D and 3D features. The evaluation of the method is twofold. First, an analysis of variance on input parameters is conducted to estimate the sensitivity of our method to their initialization. Second, for a selected set of their values, the target registration error (TRE) was calculated on 29,760 (resp. 4000) registrations in two different experiments. It was obtained using manually identified anatomical fiducials. RESULTS: For 160 US volumes, from 20 patients, recorded during routine biopsy procedures performed in two hospitals by six operators, the mean TRE was [Formula: see text] mm (resp. [Formula: see text] mm). CONCLUSION: This work allows envisioning further developments for prostate navigation and their clinical transfer.
Authors: Sheng Xu; Jochen Kruecker; Baris Turkbey; Neil Glossop; Anurag K Singh; Peter Choyke; Peter Pinto; Bradford J Wood Journal: Comput Aided Surg Date: 2008-09