Automatic Prostate Region Segmentation in MRI

Abstract

For decades now, medical imaging (MRI, CT, PET, etc) has been an indispensable aspect of the medical industry for practitioners to conduct effective diagnostic processes. An MRI, in particular, is vastly more adept at capturing detailed images than a CT or X-Ray scan. Image Segmentation is the process of partitioning an image into multiple segments that represent a similar meaning. The use of automatic image segmentation has been increasingly popular in medical imaging to identify the shape, size, and location of bone structures, organ tumors, and more.

This project focuses on the automatic segmentation of the inner and surrounding regions of the prostate in MRI. The prostate, which is located in a body's pelvis, can be partitioned into 4 anatomical zones, which include – peripheral zone (PZ), anterior fibromuscular stroma (AFS), transition zone (TZ), and central zone. The automatic and accurate segmentation of the prostate and its zone is crucial to various applications including prostate cancer assessment and tumor identification.

This project examines, assesses and implements the latest deep learning-based method for the automatic segmentation of the prostate, named TransUNET. To add, an Atrous Spatial Pyramid Pooling (ASPP) module is created which accommodates for the varying size and shape changes of the prostate regions in different MRI slices, and is successfully used to modify the TrasnUNet method. From the results of the experiments, the modified TransUNet outperforms the the original TransUNet by some ~1%.