Automated Detection and Localization of Transition Zone Prostate Cancer on Magnetic Resonance Images using Deep Learning
Year: 2021
Type: dissertation
Abstract: Given the high frequency of prostate cancer (PCa), multi-parametric magnetic resonance imaging (MP-MRI) is increasingly used to aid PCa diagnosis and reduce overdiagnosis. However, current practice for interpreting prostate MP-MRI in undiagnosed patients for PCa has limited accuracy and low inter-observer agreement in transition zone (TZ) of the prostate. I describe development of a fully automated computer-aided detection method for TZ PCa on apparent diffusion coefficient (ADC) map MR images using a deep learning segmentation model. Different U-Net models were considered for prostate and TZ segmentation, and the 3D nnU-Net had the best performance with a Dice score of 0.822 for prostate segmentation. An ensemble of three 2D U-Nets of varying hyperparameters, which used the ADC map and predicted prostate segmentation by the nnU-Net as input, achieved a sensitivity and specificity of 0.633 and 0.795, performing better than previous studies of similarly sized datasets in fully automated TZ PCa detection.
Author: Timothy Wong
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3157190864