| The most common treatment of prostate cancer (CaP) is via radical prostatectomy (RP), of which 75,000 are performed in the United States each year. However, within the current paradigm, 15-40% of RP treatments ultimately fail in the form of biochemical recurrence (BCR) within 5 years. Gleason scoring, derived from visual inspection of tissue morphology, has been the gold standard for distinguishing aggressive CaP for over 40 years. Furthermore, the current initiative towards personalized health care has attempted to utilize an integrated predictor via molecular markers such as prostate specific c antigen (PSA) to identify men with aggressive localized CaP. However, the non-specificity of these tests has led to an over-treatment of CaP, which is responsible for increased morbidity that is both stressful and costly for the patient. This dissertation attempts to develop the algorithms that could pave the way for a new class of integrated predictors, which can combine histomorphometric and molecular features into an integrated biomarker and present the information needed for better patient care. Our overall goal was to predict BCR in CaP patients following RP treatment. A host of novel machine learning tools were developed to create integrated diagnostic tests, including dimensionality reduction (Adaptive Dimensionality Reduction with Semi-Supervision (AdDReSS)) and data integration (Supervised Multi-view Canonical Correlation Analysis (sMVCCA)) methodologies to handle complex, non-linear, high dimensional and heterogeneous biomedical data. Furthermore, the development and discovery of unique discriminatory features for differentiating aggressive CaP were necessary for the understanding of cancer progression and the foundation of an integrated biomarker. Novel histomorphometric features (Co-occurring Gland Tensors (CGTs) and Cell Orientation Entropy (COrE)) were developed to quantify important differentiating image-based characteristics of CaP morphology. These methods were shown to outperform Kattan nomogram and Gleason scoring for predicting BCR following RP. Lastly, fusion of histomorphometry and protein expression into an integrated signature was performed via sMVCCA, and demonstrated improved identification of men with BCR following RP compared to histomorphometric and proteomic signatures alone. |
