MR-based radiomics-clinical nomogram in epithelial ovarian tumor prognosis prediction: tumor body texture analysis across various acquisition protocols

Tianping Wang, Haijie Wang, Yida Wang, Xuefen Liu, Lei Ling, Guofu Zhang, Guang Yang, He ZhangJournal of Ovarian Research 2022 · DOI

Abstract: MR-based radiomics analysis showed high accuracy in the prognostic estimation of epithelial ovarian cancer (EOC) patients and could help predict therapeutic outcomes before treatment.

Introduction

Epithelial ovarian cancer (EOC) is the most lethal gynecological malignancy. While many patients initially respond to surgery and chemotherapy, a staggering majority experience relapse within three years. Currently, clinicians rely on markers like CA125 levels and FIGO staging to estimate outcomes, but these often lack the precision needed for truly individualized treatment planning.

The challenge lies in tumor heterogeneity—the fact that cancer cells within a single tumor can vary significantly. This study explores whether radiomics (the high-throughput extraction of quantitative features from medical images) can capture this "hidden" heterogeneity to provide a non-invasive, pre-treatment roadmap for a patient's likely disease progression.

Methodology

The researchers analyzed a retrospective dataset of 186 pathologically proven EOC cases from the Obstetrics and Gynecology Hospital of Fudan University (Shanghai, China), spanning 2013 to 2018. The cohort was split into a training set (130) and a validation set (56).

Radiomics Workflow

To ensure a rigorous analysis, the study employed a multi-step workflow:

  • Multi-Protocol Extraction: 1,116 radiomic features (including first-order statistics, texture, and wavelet transforms) were extracted from four MRI sequences: T1-weighted (T1WI), T2-weighted (T2WI), Diffusion-Weighted (DWI), and Contrast-Enhanced T1 (CE-T1WI).
  • Dataset Balancing: An unsupervised K-means clustering algorithm was used to balance the distribution of radiomic features across the training and validation groups, ensuring they were truly comparable.
  • Feature Selection & Modeling: Pearson Correlation Coefficients and the "Relief" algorithm were used to identify the most predictive features. Support Vector Machines (SVM) were then trained to generate a "rad-score."
  • Nomogram Construction: Finally, a visual nomogram was built to integrate the MRI rad-score with clinical factors like age and FIGO stage.
Overview of the radiomics workflow: (A) Image segmentation across MRI protocols, (B) Extraction of radiomics and clinical features, and (C) Comprehensive analysis including ROC curves, DCA, and nomogram construction.

Key Results and Discussion

The study revealed several critical insights into the prognostic power of MRI-based radiomics:

  • T2WI as the Primary Predictor: Surprisingly, T2-weighted imaging—a standard sequence often used primarily for anatomical detail—yielded better prognostic information than "functional" sequences like DWI or contrast-enhanced scans.
  • Superiority of 3D Analysis: The study found that 3D analysis (whole tumor volume) provided more stable and accurate predictions than 2D analysis (single largest slice), likely because it captures the full spatial complexity of the tumor.
  • Clinical-Radiomic Synergy: Combining the T2WI rad-score with age and FIGO stage created a model significantly more powerful than clinical or radiomic data in isolation.

Model Performance

The T2WI radiomic-clinical nomogram achieved an AUC (Area Under the Curve) of 0.866 in the training cohort and 0.818 in the validation cohort. This significantly outperformed the clinical-only model, which had an AUC of approximately 0.70. Decision Curve Analysis (DCA) demonstrated that using the T2WI nomogram to guide clinical decisions provides a higher "net benefit" across a wide range of risk thresholds compared to traditional clinical markers alone.

3D visualization of the feature distribution, illustrating the high-dimensional heterogeneity captured by the radiomics approach.

Conclusion and Future Directions

While the radiomics-clinical nomogram shows great promise, the authors acknowledge several hurdles before clinical implementation. As a single-center retrospective study, the model requires external validation on data from different scanners and institutions. Additionally, the current manual segmentation process is time-consuming; the authors suggest that Deep Learning could eventually automate this process, making the tool faster and less subjective for clinical use. Refinement with larger cohorts will also be necessary to improve the model's accuracy for rarer EOC subtypes.

Summary updated

Abstract

MR-based radiomics analysis showed the high accuracy in prognostic estimation of EOC patients and could help to predict therapeutic outcome before treatment.
BibTeX
@article{wang2022mr,
  author = {Tianping Wang and Haijie Wang and Yida Wang and Xuefen Liu and Lei Ling and Guofu Zhang and Guang Yang and He Zhang},
  title = {{MR-based radiomics-clinical nomogram in epithelial ovarian tumor prognosis prediction: tumor body texture analysis across various acquisition protocols}},
  journal = {Journal of Ovarian Research},
  year = {2022},
  doi = {10.1186/s13048-021-00941-7},
}
Figure