End-to-end integrative segmentation and radiomics prognostic models improve risk stratification of high-grade serous ovarian cancer: a retrospective multi-cohort study
File(s)
Accepted version
Author(s)
Type
Journal Article
Abstract
Background
Valid stratification factors for patients with epithelial ovarian cancer (EOC) are still lacking and
individualisation of care remains an unmet need. Radiomics from routine Contrast Enhanced
Computed Tomography (CE-CT) is an emerging, highly promising approach towards more
accurate prognostic models for the better preoperative stratification of the subset of patients with
high-grade-serous histology (HGSOC). However, requirements of fine manual segmentation
limit its use. To enable its broader implementation, we developed an end-to-end model that
automates segmentation processes and prognostic evaluation algorithms in HGSOC.
Methods
We retrospectively collected and segmented 607 CE-CT scans across Europe and United States.
The development cohort comprised of patients from Hammersmith Hospital (HH) (n=211),
which was split with a ratio of 7:3 for training and validation. Data from The Cancer Imagine Archive (TCIA) (United States, n=73) and Kliniken Essen-Mitte (KEM) (Germany, n=323) were
used as test sets. We developed an automated segmentation model for primary ovarian cancer
lesions in CE-CT scans with U-Net based architectures. Radiomics data were computed from the
CE-CT scans. For overall survival (OS) prediction, combinations of 13 feature reduction
methods and 12 machine learning algorithms were developed on the radiomics data and
compared with convolutional neural network models trained on CE-CT scans. In addition, we
compared our model with a published radiomics model for HGSOC prognosis, the radiomics
prognostic vector. In the HH and TCIA cohorts, additional histological diagnosis,
transcriptomics, proteomics, and copy number alterations were collected; and correlations with
the best performing OS model were identified. Predicated probabilities of the best performing
OS model were dichotomised using k-means clustering to define high and low risk groups.
Findings
Using the combination of segmentation and radiomics as an end-to-end framework, the
prognostic model improved risk stratification of HGSOC over CA-125, residual disease, FIGO
staging and the previously reported radiomics prognostic vector. Calculated from predicted and
manual segmentations, our automated segmentation model achieves dice scores of 0.90, 0.88,
0.80 for the HH validation, TCIA test and KEM test sets, respectively. The top performing
radiomics model of OS achieved a Concordance index (C-index) of 0.66 ± 0.06 (HH validation)
0.72 ± 0.05 (TCIA), and 0.60 ± 0.01 (KEM). In a multivariable model of this radiomics model
with age, residual disease, and stage, the C-index values were 0.71 ± 0.06, 0.73 ± 0.06, 0.73 ±
0.03 for the HH validation, TCIA and KEM datasets, respectively. High risk groups were
associated with poor prognosis (OS) the Hazard Ratios (CI) were 4.81 (1.61-14.35), 6.34 (2.08-
19.34), and 1.71 (1.10 - 2.65) after adjusting for stage, age, performance status and residual
disease. We show that these risk groups are associated with and invasive phenotype involving
soluble N-ethylmaleimide sensitive fusion protein attachment receptor (SNARE) interactions in
vesicular transport and activation of Mitogen-Activated Protein Kinase (MAPK) pathways.
Valid stratification factors for patients with epithelial ovarian cancer (EOC) are still lacking and
individualisation of care remains an unmet need. Radiomics from routine Contrast Enhanced
Computed Tomography (CE-CT) is an emerging, highly promising approach towards more
accurate prognostic models for the better preoperative stratification of the subset of patients with
high-grade-serous histology (HGSOC). However, requirements of fine manual segmentation
limit its use. To enable its broader implementation, we developed an end-to-end model that
automates segmentation processes and prognostic evaluation algorithms in HGSOC.
Methods
We retrospectively collected and segmented 607 CE-CT scans across Europe and United States.
The development cohort comprised of patients from Hammersmith Hospital (HH) (n=211),
which was split with a ratio of 7:3 for training and validation. Data from The Cancer Imagine Archive (TCIA) (United States, n=73) and Kliniken Essen-Mitte (KEM) (Germany, n=323) were
used as test sets. We developed an automated segmentation model for primary ovarian cancer
lesions in CE-CT scans with U-Net based architectures. Radiomics data were computed from the
CE-CT scans. For overall survival (OS) prediction, combinations of 13 feature reduction
methods and 12 machine learning algorithms were developed on the radiomics data and
compared with convolutional neural network models trained on CE-CT scans. In addition, we
compared our model with a published radiomics model for HGSOC prognosis, the radiomics
prognostic vector. In the HH and TCIA cohorts, additional histological diagnosis,
transcriptomics, proteomics, and copy number alterations were collected; and correlations with
the best performing OS model were identified. Predicated probabilities of the best performing
OS model were dichotomised using k-means clustering to define high and low risk groups.
Findings
Using the combination of segmentation and radiomics as an end-to-end framework, the
prognostic model improved risk stratification of HGSOC over CA-125, residual disease, FIGO
staging and the previously reported radiomics prognostic vector. Calculated from predicted and
manual segmentations, our automated segmentation model achieves dice scores of 0.90, 0.88,
0.80 for the HH validation, TCIA test and KEM test sets, respectively. The top performing
radiomics model of OS achieved a Concordance index (C-index) of 0.66 ± 0.06 (HH validation)
0.72 ± 0.05 (TCIA), and 0.60 ± 0.01 (KEM). In a multivariable model of this radiomics model
with age, residual disease, and stage, the C-index values were 0.71 ± 0.06, 0.73 ± 0.06, 0.73 ±
0.03 for the HH validation, TCIA and KEM datasets, respectively. High risk groups were
associated with poor prognosis (OS) the Hazard Ratios (CI) were 4.81 (1.61-14.35), 6.34 (2.08-
19.34), and 1.71 (1.10 - 2.65) after adjusting for stage, age, performance status and residual
disease. We show that these risk groups are associated with and invasive phenotype involving
soluble N-ethylmaleimide sensitive fusion protein attachment receptor (SNARE) interactions in
vesicular transport and activation of Mitogen-Activated Protein Kinase (MAPK) pathways.
Date Acceptance
2025-11-21
Citation
The Lancet: Digital Health
ISSN
2589-7500
Publisher
Elsevier
Journal / Book Title
The Lancet: Digital Health
Copyright Statement
Copyright © 2025 Copyright Owner. This is the author’s accepted manuscript made available under a CC-BY licence in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy)
License URL
Publication Status
Accepted