Exploring Magnetic Resonance Imaging Biomarkers of Radiation Therapy Response for Advanced Cervical Cancer
Abstract
Purpose
To evaluate longitudinal changes in structural MRI biomarkers across three radiotherapy timepoints in advanced cervical cancer and their association with treatment response and recurrence.
Methods
MRI images from 23 patients were extracted from the Cancer Imaging Archive. Images were acquired at 3 timepoints: before treatment, early during treatment and mid-treatment. Shape and first order radiomics features were extracted at each timepoint using the pyRadiomics package. Delta radiomics were computed to assess changes in tumor characteristics between baseline and follow-up timepoints (Δ21 and Δ31) for each feature. Feature selection and quantitative analysis were conducted using a robust Python pipeline. Preprocessing within cross-validation folds included feature standardization and variance-based feature filtering. Model development used a nested leave-one-out cross-validation framework. Feature selection and hyperparameter tuning were performed within training data to prevent overfitting. Feature stability was assessed by repeating feature selection within training data, retaining only on features selected in more than half of repetitions. Model performance was evaluated using AUC, sensitivity, specificity, and bootstrap confidence intervals. A permutation test was conducted to assess predictability.
Results
The model for Δ21 demonstrated strong discrimination of delta radiomics with an AUC of 0.83 (95% CI: 0.615–1.000) and a permutation p-value of 0.004. The model for Δ31 also showed a good discrimination of delta radiomics with an AUC of 0.74 (95% CI: 0.500-0.931) and a p-value of 0.0270. VoxelVolume was the most stable feature, selected in all 23 cross-validation folds, and yielded a univariate AUC of 0.685 for Δ21 and 0.762 for Δ31.
Conclusion
Early tumor volume changes on T2 MRI showed good predictive value for treatment response, supporting their potential utility as early biomarkers enabling individualized treatment in advanced cervical cancer. Future work will include external validation of the model and extending the analysis to include longitudinal PET and CT imaging.