Current research in radiology field is increasingly focusing on developing computer aided detection (CAD) systems able to support radiologists in the detection of suspicious regions, reducing oversight, errors and working time. Prostate cancer (PCa) is the most common cancer afflicting men in USA. Multiparametric Magnetic Resonance (mp-MR) imaging is recently emerging as a powerful tool for PCa diagnosis. The development of CAD systems for its automatic processing and elaboration is growing but they can be affected by the variation of the imaging characteristics of PCa depending on its aggressiveness and location. The aim of this study is to characterize the homogeneity of a large set of data derived from mp-MR images, in order to assess the effect on the performances of a CAD system for PCa detection. Firstly, 15 semiquantitative and quantitative features were extracted from malignant and normal region of interest in 60 patients, who underwent mp-MR exam before prostatectomy. Then, we used a clustering procedure based on a Self-Organizing Map (SOM) for grouping patients with similar characteristics from the features point of view. Finally, we evaluated the impact of this partition on the malignant voxel detection by means of a classifier based on a set of SOMs trained and tested using only those patient belonging to the same cluster. We compared these results with those obtained using a unique classifier for all patients. From our analysis it emerged that the image partition in homogeneous groups can effectively improve the final detection performances.
Download Full PDF Version (Non-Commercial Use)