Viswanath, Satish Easwar. A quantitative data representation framework for structural and functional MR Imaging with application to prostate cancer detection. Retrieved from https://doi.org/doi:10.7282/T3HQ3XVH
DescriptionProstate cancer (CaP) is currently the second leading cause of cancer-related deaths in the United States among men, but there is a paucity of non-invasive image-based information for CaP detection and staging in vivo. Studies have shown the utility of multi-protocol magnetic resonance imaging (MRI) to improve CaP detection accuracy by using both T2-weighted (T2w), dynamic contrast enhanced (DCE), and diffusion weighted (DWI) MRI information. In this thesis, we present methods for quantitative representation of structural and functional imaging data with the objective of building automated classifiers to improve CaP detection accuracy in vivo. In vivo disease presence was quantified via extraction of textural signatures from T2w MRI. Evaluation of these signatures showed that CaP appearance within each of the two dominant prostate regions (central gland, peripheral zone) is significantly different. A classifier trained on zone-specific features also yielded a higher detection accuracy compared to a simpler, monolithic combination of all the texture features. While a number of automated classifiers are available, classifier choice must account for limitations in dataset size and annotation (such as with in vivo prostate MRI). A comprehensive evaluation of different classifier schemes was undertaken for the specific problem of automated CaP detection via T2w MRI on a zonewise basis. It was found that simple classifiers yielded significantly improved CaP detection accuracies compared to complex classifiers. Fundamental differences must be overcome when constructing a unified quantitative representation of structural (T2w) and functional (DCE, DWI) MRI. We present a novel technique, referred to as consensus embedding, which constructs a lower dimensional representation (embedding) from a high dimensional feature space such that information (class-based or otherwise) is optimally preserved. Consensus embedding is shown to result in an improved representation of the data compared to alternative DR-based strategies in a variety of experimental domains. A unified quantitative representation of T2w, DCE, and DWI prostate MRI was constructed via the consensus embedding framework. This yielded an integrated classifier which was more accurate for CaP detection in vivo as compared to using structural and functional information individually, or using a naive combination of such differing types of information.