Texture Features and Artificial Neural Networks: A Way to Improve the Specificity of a CAD System for Multiparametric MR Prostate Cancer

Abstract

Prostate cancer (PCa) is the most common cancer afflicting men in USA. Multiparametric Magnetic Resonance imaging is recently emerging as a powerful tool for PCa diagnosis, but its analysis and interpretation is time-consuming and affected by the radiologist experience. For this reason, current research is focusing on developing computer aided detection (CAD) systems able to support radiologists in the PCa diagnosis. Although several studies proposed CAD systems with very high performances in terms of sensitivity, the analysis of false positive (FP) areas is usually not clearly presented. In this study we propose a new method for the reduction of FP voxels based on the analysis of the textural information contained in the T2 weighted images and the apparent diffusion coefficient maps. In this method the 64 textural features are first discretized and selected to reduce the data variability and remove irrelevant variables, then fed into an Artificial Neural Network able to distinguish between malignant and healthy areas. In this study we apply the method to a previously developed CAD system, and results show a significant decrease of the number of FP voxels with respect to the CAD system and an increase of the precision of PCa segmentation. Having less FP and more precise PCa segmentation areas, could contribute to develop CAD system able to provide PCa characterization, which represents the key to personalized treatment options.