Abstract:
Worldwide, prostate cancer is the second most prevalent cancer among men. Currently, invasive
prostate biopsies are the only method of diagnosing prostate cancer. To provide a better and
less invasive diagnostic method, a Raman Spectroscopy probe has recently been developed at
the University of Auckland. The probe has shown its ability to differentiate between malignant
and benign prostate tissue and identify the Gleason Grade of the tumour. Recent literature
has shown artificial intelligence (AI) based models have promise for automatically segmenting
prostate tumours. This thesis aimed to apply AI-based models to identify tumours on multiparametric magnetic resonance imaging (mpMRI) to help guide where the Raman Spectroscopy
probe should acquire measurements within the prostate. To achieve this, three methods were
applied to segment tumours. These methods utilised T2-weighted (T2W), apparent diffusion
coefficient (ADC) maps from diffusion weighted imaging (DWI), and a high b-value DWI image
from three different prostate MRI datasets. The first method involved training an open-source
deep learning model, nnU-Net. An iterative methodology was applied to train four different
nnU-Net models, differing in their inputs and outputs. After the iterative development process,
the best nnU-Net model achieved a dice similarity coefficient, sensitivity, and specificity of 0.552,
0.509, and 0.997, respectively. The second method involved running a pre-trained convolutional
neural network. This model achieved a dice similarity coefficient, sensitivity, and specificity of
0.176, 0.818, and 0.801, respectively. The final method involved thresholding the ADC maps.
This model achieved a dice similarity coefficient, sensitivity, and specificity of 0.372, 0.449, and
0.987, respectively. Therefore, the best tumour segmentation method investigated in this thesis
was the nnU-Net model. In addition to segmenting the tumour, a machine learning model was
developed to determine if a region where a biopsy was completed was likely to contain tumour.
To complete this, a random forest classifier was trained with mpMRI radiomic features that were
extracted from biopsy regions. This classifier achieved an accuracy, sensitivity, and specificity
of 0.664, 0.652, and 0.667 respectively. Both the classifier and tumour segmentation methods
provide an important first step towards developing AI-based models to help guide the Raman
Spectroscopy probe.