3740 - Data-Driven Pre-Implant AI Predictions for Supplemental Interstitial Needle Needs In Hybrid Cervix Brachytherapy

Purpose/Objective(s): Hybrid intracavitary/interstitial (IC/ISBT) brachytherapy provides dosimetric advantages in treating cervical cancer; however, pre-procedure estimation of interstitial needle usage remains largely experience-driven. We developed an AI-based machine learning framework that leverages quantitative geometric descriptors derived from pre-planning anatomy to identify patients likely to require more extensive interstitial needle utilization.

Materials/Methods: Seventy-five stage I–IV cervix patients treated definitively with hybrid tandem-and-ovoid HDR brachytherapy between 2022–2025 were retrospectively analyzed. Interstitial needle utilization was defined as the relative dose contribution from supplemental needles to the total delivered dose (range 2.5–70.4%, mean±SD 19.7±16.5%). Cases were stratified using a 15% threshold to distinguish moderate (=15%, n=37) versus high (>15%, n=38) needle contribution. For each case, 196 quantitative geometric variables were computed, including shape descriptors of the high-risk clinical target volume (HR-CTV) and organs at risk (OARs) as well as spatial relationship metrics among targets, OARs, and applicator geometry. Ten-fold cross-validation was performed. To reduce overfitting, feature selection was conducted within each fold using mutual-information screening followed by redundancy-aware selection to minimize collinearity. Gradient boosting, logistic regression, and a pairwise ranking model incorporating continuous needle contribution were trained using 30 selected features with sample weighting reflecting adherence to EMBRACE planning recommendations. Model outputs were integrated using a constrained soft-voting ensemble with optimized weights. Feature importance was evaluated using cross-validated permutation analysis.

Results: The final model achieved an AUC of 0.95, accuracy of 0.91, sensitivity of 0.93, specificity of 0.87, and F1 score of 0.88 for identifying cases requiring high needle utilization. Performance exceeded models using FIGO stage and HR-CTV volume alone (AUC 0.81, accuracy 0.81). The most predictive feature was the maximum superior-inferior extent of the HR-CTV relative to the ovoid plane. Additional influential predictors included HR-CTV shape characteristics (axis metrics, elongation, and sphericity) and OAR geometric descriptors (rectal minimum 3D diameter, bladder axis metrics, bowel surface area and elongation), highlighting the combined impact of tumor morphology and adjacent anatomy.

Conclusion: A data-driven machine learning approach can prospectively identify cervical brachytherapy cases likely to require greater interstitial needle usage. The model likely outperforms traditional clinical parameters because it captures more comprehensive spatial relationships between the target and surrounding OARs. This framework demonstrates feasibility for decision support to guide implant strategy and improve consistency in hybrid IC/ISBT.