Textural Analysis of the Prostate using Co-occurence matrices
A. Knowles, Oscar Brihuega‐Moreno, Peter Gibbs, Lindsay W. Turnbull · 2001
Introduction Magnetic resonance (MR) imaging is currently unable to reliably differentiate prostatic adenocarcinoma from benign prostatic hypertrophy (BPH) due to an overlap in the appearance of these two pathologies on different MR sequences1. To improve differentiation dynamic contrast-enhanced imaging2 can also be used but there still exists room for improvement in diagnostic accuracy. Previous work carried out within our centre has attempted to classify normal, benign and malignant disease within the prostate by neural network analysis of the DCE-MRI3. The resultant neural network predicted the biopsy result with an accuracy of 89%. We have also attempted to use 2D neural networks to look at 'look' at selected regions giving a resulting accuracy of 83%4. In this study we investigate the use of co-occurrence matrices5 in order to quantify the spatial variations in grey tone which exist. The textural descriptors derived are then used to train a back-propagation neural network6 which is then used for classification. Methods MR imaging was performed on 10 patients using a 1.5 T Signa Advantage (General Electric Medical Systems, Milwaukee) using a pelvic phased array coil for signal reception. Initial localising images were acquired in the sagittal plane to ensure correct positioning of the pelvic phased array coil. A series of axial T2 weighted Fast Spin Echo (FSE) images (TR/TE = 12750/130 ms, 20 cm field of view) with a slice thickness of 2.5cm, matrix of 256*192 were acquired from the apex to the base of the prostate (see Image 1).