UCLA

In this work, we aimed to develop methods and approaches to enable the use of unannotated or weakly annotated clinically generated datasets in clinical data science and deep learning, in the clinical context of prostate cancer and multiparametric MRI of the prostate. Specifically, we demonstrate: 1) The development of an optimized regional targeted biopsy strategy that could reduce the number of biopsies that need to retrieved in a targeted biopsy procedure, by creating a combined MRI, ultrasound, and histopathological evaluation dataset from the clinical record, 2) the creation of a state-of-the-art prostate organ segmentation model using unrefined clinically-generated annotations as well as an evaluation of the utility of those annotations to improve model training on small strongly annotated datasets, 3) the training of a high performance segmentation model on private data originating from three different healthcare institutions using the federated learning approach, without requiring any data to be transferred across institutional boundaries, and 4) the creation of patient-level predictive models for prostate cancer risk stratification from multiparametric MRI of the prostate, and an evaluation of the relative contribution of pretrained voxel-level feature extractors using unannotated, weakly annotated, and strongly annotated data with the finding that even an unannotated data-based pretrained model is effective. The contributions of this dissertation demonstrate the potential uses of unannotated and weakly annotated clinically generated data in clinical data science and machine learning model development for healthcare, and enable the development of clinical tools for the prostate cancer clinical workflow.