Machine Learning-Based Prognostic Modelling Using MRI Radiomic Data in Cervical Cancer Treated with Definitive Chemoradiotherapy and Brachytherapy

TL;DR

This study shows that combining clinical data with MRI-based radiomic features improves survival prediction in cervical cancer patients undergoing chemoradiotherapy and brachytherapy.

Contribution

The study introduces a novel application of CatBoost machine learning with radiomic features for prognosis prediction in locally advanced cervical cancer.

Findings

Models combining clinical and radiomic features outperformed clinical-only models in accuracy and F1-score.

Radiomic features from T1W and T2W MRI sequences significantly enhanced predictive performance.

The CatBoost_CLI + T2W_DMFS model achieved 92.31% test accuracy for distant metastasis-free survival prediction.

Abstract

Locally advanced cervical cancer is commonly treated with chemoradiotherapy and 3D image-guided adaptive brachytherapy (3D-IGABT). While advances in systemic therapies and radiotherapy techniques have improved survival and reduced side effects, the disease remains prevalent in low-resource settings, making accurate pretreatment prognosis increasingly important. This study evaluated CatBoost-based machine learning models for survival prediction in patients with locally advanced cervical cancer. Results showed that models integrating both clinical and radiomic features outperformed those using only clinical data, with notable improvements in accuracy and F1-score. Radiomic features, particularly from T1-weighted (T1W) and T2-weighted (T2W) MRI sequences, significantly enhanced the models’ predictive performance. The study stands out for its focus on a relatively under-researched cancer stage, its use of long-term follow-up data, and its comprehensive inclusion of patient and treatment characteristics. These findings contribute valuable insight into prognosis prediction in settings where cervical cancer is most prevalent.

Background: This study aims to evaluate the contribution of clinical and radiomic features to machine learning-based models for survival prediction in patients with locally advanced cervical cancer. Methods: Clinical and radiomic data from 161 patients were retrospectively collected from a single center. Radiomic features were obtained from contrast-enhanced magnetic resonance imaging (MRI) T1-weighted (T1W), T2-weighted (T2W), and diffusion-weighted (DWI) sequences. After data cleaning, feature engineering, and scaling, survival prediction models were created using the CatBoost algorithm with different data combinations (clinical, clinical + T1W, clinical + T2W, clinical + DWI). The performance of the models was evaluated using test accuracy, precision, recall, F1-score, ROC curve, and Bland–Altman analysis. Results: Models using both clinical and radiomic features showed significant improvements in accuracy and F1-score compared to models based solely on clinical data. In particular, the CatBoost_CLI + T2W_DMFS model achieved the best performance, with a test accuracy of 92.31% and an F1-score of 88.62 for distant metastasis-free survival prediction. ROC and Bland–Altman analyses further demonstrated that this model has high discriminative power and prediction consistency. Conclusions: The CatBoost algorithm shows high accuracy and reliability for survival prediction in locally advanced cervical cancer when clinical and radiomic features are combined. The addition of radiomics data significantly improves model performance.

Figures