Prostate Image Segmentation Using Deep Learning

Abstract

Biomedical imaging and analysis are recognised as potential non-invasive methods for identifying various life-threatening conditions. The evolution and widespread availability of magnetic resonance imaging (MRI) technology and artificial intelligence (AI) have paved the way for computer aided detection and diagnosis of these diseases, particularly with multiparametric MRI modalities such as T2-weighted (T2W) MRI, dynamic contrast-enhanced MRI and diffusion-weighted MRI. These advancements open up new avenues for significantly enhancing diagnostic capabilities through segmentation of vital organs. This thesis focuses on segmentation of the prostate gland and membranous urethra (MU) to inform studies on prostate cancer, urinary incontinence and benign prostatic hyperplasia. Three studies that exploit the availability of multisite, multiparametric, and multiplanar T2W for enhancing prostate segmentation accuracy are presented. The first study evaluates prostate segmentation performance of the proposed deep learning model 3dDOSPyUSENet from multisite heterogeneous T2W MRI scans and concludes that pyramid pooling, deep output supervision, and squeeze and excitation modules contribute to domain generalisation. Also, the incorporation of a domain invariant surface attention based training strategy led to enhanced segmentation accuracy and domain generalisability. In the second study, a novel multi-encoder single-decoder deep neural network architecture named 3dDOSPyResidualUSENet is proposed, with a focus on extracting both anatomical and functional features from multiparametric and multiplanar images for prostate segmentation. The findings concluded that combining multiple planes of T2W images, and multiple MRI modalities, as well as multiple planes with multiple modalities, resulted in improved segmentation performance. The focus of the third study is on MU segmentation, aimed at mitigating urinary incontinence and anticipated treatment-related morbidities following radical prostatectomy. A traditional image processing pipeline is proposed combining Hough transform and active contours, while cross-referencing multiplanar MRI and sparse markup annotations. A centreline Dice based deep learning training strategy outperforms the traditional image processing pipeline given the availability of pixelwise MU annotations. The technical contributions hold promise for enhancing segmentation and reducing the effort required for manual interpretation of biomedical imaging, thereby enhancing radiologist efficiency. Moreover, this thesis offers insights into potential avenues for further development of the presented methodologies to enhance their efficacy and applicability in future endeavours.