MRI Confirmed Prostate Tissue Classification with Laplacian Eigenmaps of Ultrasound RF Spectra

The delivery of therapeutic prostate interventions can be improved by intraprocedural visualization of the tumor during ultrasound-guided procedures. To this end, ultrasound-based tissue classification and registration of the clinical target volume from preoperative multiparametric MR images to intraoperative ultrasound are suggested as two potential solutions. In this paper we report techniques to implement both of these solutions. In ultrasound-based tissue typing, we employ Laplacian eigenmaps for reducing the dimensionality of the spectral feature space formed by ultrasound RF power spectra. This is followed by support vector machine classification for separating cancer from normal prostate tissue. A classification accuracy of 78.3±4.8% is reported. We also present a deformable MR-US registration method which relies on transforming the binary label maps acquired by delineating the prostate gland in both MRI and ultrasound. This method is developed to transfer the diagnostic references from MRI to US for training and validation of the proposed ultrasound-based prostate tissue classification technique. It yields a target registration error of 3.5±2.1 mm. We also report its use for MR-based dose boosting during ultrasound-guided brachytherapy.