Predicting the Need for Supplemental Interstitial Needles In Hybrid Intracavitary/Interstitial (IC/IS) High-Dose-Rate Cervical Brachytherapy
Abstract
Purpose
Recent studies have demonstrated the benefit of incorporating supplemental needles into intracavitary brachytherapy. However, estimating the extent of interstitial needle usage needed prior to the procedure remains challenging. This study aimed to develop a machine learning framework to predict the need for supplemental needles using pre-planning geometric information.
Methods
Since 2019, our institution has routinely added supplemental needles using 3D-printed tandem-anchored interstitial templates for all tandem-and-ovoid cases. Seventy-five hybrid intracavitary/interstitial brachytherapy (IC/ISBT) patients treated between 2022 and 2025 were retrospectively analyzed. For each patient, the final HDR fraction was evaluated. A total of 196 geometric features (165 shape-based and 31 distance-based) describing the HR-CTV, organs at risk (bladder, rectum, and bowel), and their spatial relationships to the tandem were extracted. The percentage needle contribution to the total dose quantified interstitial contribution (range: 2.5%–70.4%, mean±SD: 19.7±16.5%). A cutoff of 15% defined moderate (n=37, mean needles: 2.5) versus high (n=38, mean needles: 4.7) needle contribution. Three complementary models (gradient boosting, logistic regression, pairwise ranking) were trained using 10-fold cross-validation. The ranking model incorporated ordinal learning loss from continuous needle contribution to inform binary clinical decisions. Training cases were weighted according to plan quality relative to EMBRACE trial aims. Model predictions were combined using constrained soft-voting ensemble with optimized weights. Performance was evaluated using ROC and confusion matrices analysis. Feature importance was assessed using cross-validated logistic regression coefficients.
Results
The final ensemble achieved an AUC of 0.82, accuracy of 0.76, specificity of 0.81, sensitivity of 0.71, and F1 score of 0.75 in predicting moderate/high needle need, compared with an AUC of 0.33 and accuracy of 0.51 using FIGO stage alone.
Conclusion
This framework demonstrates the feasibility of predicting the need for supplemental needles in IC/ISBT using geometric features, supporting prospective implant decisions (e.g. estimate needle counts) and improving procedural consistency.