[Computer-aided analysis of transrectal ultrasound images of the prostate].

INTRODUCTION Prostate cancer is usually diagnosed by transrectal ultrasound (TRUS) biopsy. Nevertheless, suspicious images are frequently not found. Imaging analysis studies aim to identify ultrasound patterns characteristic of apparently hidden conditions. MATERIAL AND METHODS We digitally recorded 288 TRUS ultrasound guided transrectal biopsies and extracted 3 static images from the puncture-biopsy area. The extraction of the texture characteristics were obtained by "simple mapping" on a gray scale and spatial gray level dependence matrices (SGLDM), also known as Haralick's co-occurrence matrices, which study the relationship of each pixel and its neighbors. A pattern recognition software system was developed with two different classification methods: nearest neighbor (k-NN) and Markov's hidden models. Finally, a virtual experiment was carried out in which four urologists compared their diagnostic accuracy for prostate cancer with our system in 408 TRUS images, not in real time. RESULTS The diagnostic capacity (R.O.C. curve) with the simple gray map study was 59.7% with nearest-neighbor classification and 61.6% with Markov's hidden models classification. The co-occurrence matrices showed an area under R.O.C. curve of 60.1% and 60.0% with k-NN and Markov's hidden models classification, respectively. The virtual experiment was conducted with a simple gray map study and k-NN classification. The images processed by our system showed the following diagnostic accuracy: 63.3, 67, 64.3 and 63.7% compared to 61.7, 60.5, 66.2 and 60.7% with the original image. CONCLUSIONS Our pattern recognition system for prostate cancer TRUS images has a limited, yet stable, accuracy.