Prostate cancer is the most commonly diagnosed cancer and the third highest cause of cancer-related mortality in men. Treatment for prostate cancer is quite successful, with about a 95% 5 year survival rate for patients with cancer stage below 3. However, this success hinges on an early stage diagnosis and confirmation. While it is imperative to build a powerful predictive model for prostate cancer imaging data, existing methods cannot be applied due to their inadequacy of accommodating the unique features of prostate cancer imaging data. In particular, data imbalance, spatial correlation, and outcome misclassification present great challenges in data analysis. In this talk, I will discuss various statistical approaches to building an effective prediction model. I will examine the data from multiple angles with their features accommodated differently.