Prostate Lesion Detection and Segmentation Using Convolutional Neural Networks

Date of Award

7-1-2022

Document Type

Thesis

Degree Name

Master of Science in Data Science

First Advisor

Patricia Angela R. Abu, PhD

Abstract

Prostate cancer is the 3rd most diagnosed cancer overall. Some screening methods result in overdiagonosis and overtreatment while others are invasive. Medical advancements have allowed the use of multiparametric MRI — a noninvasive and reliable screening process for prostate cancer. However, assessment would vary from different professionals introducing subjectivity. While convolutional neural networks have been used in multiple studies to objectively segment prostate lesions, due to the varying ground-truth established, it is not possible to reproduce and validate the results. This study involved executing a repeatable framework for detecting presence and segmenting prostate cancer lesions using annotated apparent diffusion coefficient maps from a publicly available dataset that includes multiparametric MRI images of 15 patients that are confirmed or suspected of prostate cancer with two studies each. For classification, the study used Inception V3 with transfer learning trained with RandAugment and dropout magnitudes of 0 and 0.2. For segmentation, the study used U-Net with batch normalization tested with different encoders. Data image augmentation combinations and hyperparameters adopted from published frameworks were implemented to validate which combination of parameters works best. The resulting framework was able to achieve a classification accuracy, F1 score, precision, recall of 0.74, 0.59, 0.53, 0.67 respectively and a Dice score of 0.47 (0.44-0.49).

Share

COinS