Despite its steady reduction in mortality due to early detection and improved treatment, prostate cancer remains the most common cancer form in men in the developed countries. Multi-parametric magnetic resonance imaging is gaining clinical relevance and is increasingly used to diagnose prostate cancer. In the last 15 years, computer-aided detection systems that aid the radiologists in their clinical decision making have come into focus of medical image analysis. These frameworks normally detect cancer by computing pixel-based features or compute region-based features and give a diagnosis about a region of interest that was manually annotated. In this thesis, we propose a computer-aided detection system that automatically segments the prostate into specific regions of interest without the need for manual annotation. By incorporating a multi-modal, superpixelbased oversegmentation of the prostate into our framework, more accurate region-based features can be calculated. The system is evaluated on two datasets. The first dataset consists of multi-modal MRI scans of 20 patients of which 18 have biopsy-proven prostate cancer. The second dataset has multi-modal MRI scans of 25 patients. In both datasets, the prostate boundary, prostate zones and cancer lesions were annotated by experienced radiologists. Performance evaluation is based on receiver operating characteristic curve. The average area under the curve is 0.84 with a standard deviation of 0.08 for the first dataset. The second dataset shows an average area under the curve of 0.71 with a standard deviation of 0.11. The framework shows a better performance than comparable computer-aided detection systems in literature and proves that superpixels can improve the classification result for detecting prostate cancer (from 0.85 to 0.87 for the first and from 0.55 to 0.59 for the second dataset).