Applied machine learning strategies for radiographic image-based biomarker discovery and interpretability in ovarian cancer
File(s)
Author(s)
Linton-Reid, Kristofer
Type
Thesis
Abstract
This research adopts a hierarchical approach to investigating ovarian cancer, focusing on non-invasive radiographic imaging for screening, diagnosis, and prognosis. It analyses several unique datasets using statistical, machine learning, and deep learning techniques across three principal domains: automated segmentation, radiomics, and outcome prediction. The research begins with ovarian cancer screening, employing ultrasound data from the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS), which included 871 benign and 211 malignant cases. Here, the DeepLab model achieved Dice scores of 0.98 ± 0.032, 0.98 ± 0.014, 0.98 ± 0.018, and the Ovarian Screening Score showed F1-scores of 0.99 in training, 1.00 in validation, and 0.72 in testing. Diagnostic efforts used data from 577 masses (1444 images) from Imperial College London and 184 masses (476 images) from Morgagni-Pierantoni Hospital. The Ovarian Diagnostic Score developed surpassed the CA-125 marker, with F1 scores of 0.94 versus 0.36 in ICH validation and 0.83 versus 0.38 in MPH testing. The study also addressed prognosis in high-grade serous ovarian cancer, analysing pre-operative CT scans from Hammersmith Hospital, The Cancer Imaging Archive, and Kliniken Essen-Mitte. Using nn-Unet models, Dice scores were 0.96 (HH training), 0.90 (HH validation), 0.88 (KEM test), and 0.80 (TCIA test), with Concordance Index values of 0.85 ± 0.01 (HH training), 0.66 ± 0.06 (HH validation), 0.72 ± 0.05 (TCIA), and 0.60 ± 0.01 (KEM test). Furthermore, the research explored prognosis prediction in ovarian cancer relapse using CT scans from AstraZeneca's ICEBERG3 trial and Hammersmith Hospital data, achieving a prognostic concordance index of 0.601 ± 0.068 on test data. The study concludes with a discussion on the integration of machine learning with radiographic and molecular analyses, highlighting significant progress and areas ripe for future improvement in ovarian cancer.
Version
Open Access
Date Issued
2023-12-02
Date Awarded
01/05/2024
License URL
Advisor
Aboagye, Eric
Posma, Joram
Sponsor
Stratigrad
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)