This item is under embargo and not available online per the author's request. If you have questions, please contact digital@wumail.wustl.edu.

Date of Award

Spring 5-12-2021

Author's School

School of Engineering & Applied Science

Author's Program

Computer Science

Degree Name

Bachelor of Science

Restricted/Unrestricted

Restricted

Abstract

Prostate cancer (PCa) is among the top three most common cancers in men and has the second highest mortality rate. However, over-treatment of low grade PCas can be more harmful than just keeping it on surveillance. Hence, determination of PCa aggressiveness is crucial, but is difficult without biopsies due to convoluted signals from non-invasive methods. While DBSI metrics combined with deep neural networks (DNN) have been effective in differentiating PCa tissues, they lack in accuracy for Grade Group classification without using majority voting schema. In this study, a density-based medoid clustering model is used to detect hidden propensities in the isotropic diffusion spectrum between various PCa grades. Representative curves for each cluster were then used to predict PCa grades of other lesions to improve the diagnostic potential of prediction models. A modular neural network (MNN) was developed to combine cluster grading with the deep neural network to further increase diagnostic capabilities. Cluster grading proved to be effective, with accuracies of 76.8%, 71.2%, 81.2%, 83.3% and 87.3% for Grade Groups 1, 2, 3, 4 and 5 respectively. The developed MNN highly favors the Cluster grading model, suggesting that further improvement of DNNs is critical to increasing MNN potential.

Mentor

Dennis Barbour

Additional Advisors

Ron Cytron, Jonathan Silva

Available for download on Friday, May 12, 2023

Share

COinS