| dc.description.abstract | Prostate cancer (PCa) is one of the most prevalent cancers worldwide. When diagnosing
PCa, magnetic resonance imaging (MRI) is a useful, non-invasive tool for imaging the
prostate. Prostate MRI images are segmented to help guide a biopsy, and to determine if
a patient has PCa. When done manually, segmentation is a tedious task. For this reason,
automatic systems have been developed utilizing deep learning and convolutional neural
networks (CNNs). However, most of these systems only take the axial plane into account
when predicting the segmentation mask, even though other planar information is also
collected at the time of acquisition.
In this thesis we aim to exploit the axial, sagittal and coronal planes in order to create a
CNN for segmentation of the whole gland from prostate MRI images, based exclusively on
axial ground truth masks. We create a 2D U-net, a CNN based encoder-decoder network
and test different methodologies and model structures, for both single and multi-planar
segmentation. We also explore contrastive leaning approaches for improving performance
in a multi-planar segmentation task.
The results of the tests show that a multi-planar approach initially holds the advantage.
But, as the models get more complex with the addition of more layers, techniques and
regularizers, the axial only approach catches up. In the end both approaches reach a
dice similarity coefficient (DSC) of 91.7%, with the rest of the evaluation metrics such as
relative volume and 95% Hausdorff distance being close as well. Our results indicate that
it is possible to achieve better results by incorporating the additional planes. Nevertheless,
in our approach, this benefit was overshadowed by other additions like dropout, transfer
learning and augmentation leading to equal performance between the two models. Our
work is currently being improved and under consideration for publication. | |