Fully Automated Segmentation of High Grade Serous Ovarian Cancer on Computed Tomography Images using Deep Learning

In this thesis, we investigate how deep neural networks can be used for the fully automated segmentation of high grade serous ovarian cancer. The recent rise of deep learning has pushed the limits of what algorithms can achieve in fields of image analysis, such as the task of segmentation. The field of medical image segmentation has fundamentally changed, and it is now feasible to perform automated segmentation of complex and metastatic diseases, such as ovarian cancer. This thesis is the first work to study deep learning-based solutions for the fully automated segmentation of high grade serous ovarian cancer. The aim of the thesis is to determine whether deep learning-based segmentation for high grade serous ovarian cancer is feasible and to maximise the performance of state-of-the-art segmentation methods in this task. For this we will use a total of 451 manually labelled pre-treatment and post neoadjuvant chemotherapy computed tomography scans. First, we will use the well-established nnU-Net framework [1, 2] that automatically adapts to new datasets and segmentation problems, and offers state-of-the-art performance. For the main two disease sites, the performance of the model will be tested extensively and compared to the work of a trainee radiologist. Next, we will introduce custom methods that outperform nnU-Net in this particular segmentation task and discuss different evaluation methods for automated segmentations including mathematical distance functions and clinical assessments. Thereafter, we will consider the segmentation of the full disease burden, including very small and rare lesions. For this, different multi-class segmentation approaches are compared, and hyper-parameter tuning is applied to maximise their performance. Finally, we demonstrate how the discussed algorithms can be deployed in clinical workflows for the purpose of reducing manual annotation time. Our key results are the following. We find that deep learning-based methods can reach a performance comparable to a trainee radiologist. We suggest ways to tune the model hyper-parameters of a deep learning-based segmentation model to improve the performance over the well-established nnU-Net framework [1, 2]. Further, including classes of rare and small lesions in the approach, we find that training multiple independent models, where each only considers a subset of all classes, performs best amongst multiple different approaches. We conclude that fully automated segmentation of high grade serous ovarian cancer using deep learning-based methods is feasible using only a few hundred scans for training despite the difficulty of the segmentation problem. While the main disease sites in the pelvis/ovaries and the omentum can already be automatically segmented with high precision, some of the rare and small disease sites lack in performance. To improve the performance and ultimately the clinical value of these approaches, more effort is needed in terms of model development and creation of larger datasets.

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2022

Author dc:creator

• Buddenkotte, Thomas

Advisor dc:contributor.advisor

• Schönlieb, Carola Bibiane

• High grade serous ovarian cancer
• Deep learning
• segmentation
• hyper-paramter tuning
• Visual Turing test
• Metastatic cancer
• NVIDIA Clara

Licence

• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.87940

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/340501