Doctor of Philosophy
Hoover, Douglas A.
London Health Sciences Centre and Western University
Minimally invasive procedures for prostate cancer diagnosis and treatment, including biopsy and brachytherapy, rely on medical imaging such as two-dimensional (2D) and three-dimensional (3D) transrectal ultrasound (TRUS) and magnetic resonance imaging (MRI) for critical tasks such as target definition and diagnosis, treatment guidance, and treatment planning. Use of these imaging modalities introduces challenges including time-consuming manual prostate segmentation, poor needle tip visualization, and variable MR-US cognitive fusion. The objective of this thesis was to develop, validate, and implement software- and hardware-based tools specifically designed for minimally invasive prostate cancer procedures to overcome these challenges.
First, a deep learning-based automatic 3D TRUS prostate segmentation algorithm was developed and evaluated using a diverse dataset of clinical images acquired during prostate biopsy and brachytherapy procedures. The algorithm significantly outperformed state-of-the-art fully 3D CNNs trained using the same dataset while a segmentation time of 0.62 s demonstrated a significant reduction compared to manual segmentation. Next, the impact of dataset size, image quality, and image type on segmentation performance using this algorithm was examined. Using smaller training datasets, segmentation accuracy was shown to plateau with as little as 1000 training images, supporting the use of deep learning approaches even when data is scarce. The development of an image quality grading scale specific to 3D TRUS images will allow for easier comparison between algorithms trained using different datasets. Third, a power Doppler (PD) US-based needle tip localization method was developed and validated in both phantom and clinical cases, demonstrating reduced tip error and variation for obstructed needles compared to conventional US. Finally, a surface-based MRI-3D TRUS deformable image registration algorithm was developed and implemented clinically, demonstrating improved registration accuracy compared to manual rigid registration and reduced variation compared to the current clinical standard of physician cognitive fusion. These generalizable and easy-to-implement tools have the potential to improve workflow efficiency and accuracy for minimally invasive prostate procedures.
Summary for Lay Audience
Prostate cancer is the most common non-skin cancer in Canadian men. The management of prostate cancer often includes the use of needles, such as in diagnosis to extract tissue samples and test for the presence of cancer, known as biopsy, and in treatment to deliver radiation from inside the body, known as brachytherapy. These procedures rely on medical imaging for critical tasks including target definition, creating patient-specific treatment plans, and accurately guiding needles into the body. This also introduces challenges as time-consuming and difficult manual tasks must be completed in the operating room such as accurately identifying the prostate and needle tip locations and mentally combining information from multiple imaging types. This thesis is focused on developing innovative software and hardware solutions to overcome these challenges and improve treatment efficiency and accuracy.
First, artificial intelligence was used to train an algorithm to locate the prostate boundary (or ‘segment’) in three-dimensional ultrasound images in under 1 second, demonstrating equal segmentation accuracy and greatly reducing time compared to manual segmentation, which can take up to 10 minutes. Next, this same algorithm was tested using smaller datasets, demonstrating equivalent performance with as little as 7% of the full dataset, potentially increasing access to artificial intelligence methods even if data is scarce. Third, an alternative ultrasound technique known as power Doppler ultrasound was used to improve needle tip visibility during the live brachytherapy procedure, demonstrating reduced variability compared to standard ultrasound. Finally, an automated image registration algorithm was developed to overlay magnetic resonance images on ultrasound images, facilitating the targeting of the previously invisible tumour in the operating room.
Orlando, Nathan, "Software and Hardware-based Tools for Improving Ultrasound Guided Prostate Brachytherapy" (2022). Electronic Thesis and Dissertation Repository. 8702.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License