2722 - The Value of Pre-Brachytherapy MRI in Aiding Applicator Selection for Cervical Cancer Treatment
Presenter(s)
J. S. Chiang1, A. M. Conteh1, S. Richter2, T. Niedermayr1, and E. A. Kidd2; 1Department of Radiation Oncology, Stanford University, Stanford, CA, 2Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA
Purpose/Objective(s): Brachytherapy for gynecologic cancers often relies on intraoperative applicator selection. While precise applicator placement is critical for achieving optical outcomes, minimizing procedure duration decreases the risk of complications and patient discomfort. Pre-brachytherapy MRI (pre-MRI) may offer predictive insights. This study investigated the relationship between pre-MRI vaginal canal (VC) volumes and applicator selection and developed/evaluated machine learning models predicting applicator type based on pre-MRI features. We hypothesized that the pre-MRI upper VC volume is a significant predictor of applicator type and can be used to improve procedural planning.
Materials/Methods: One hundred and ten consecutive cervical cancer patients treated with brachytherapy at one institution (2022-2025) were retrospectively analyzed. Demographics and treatment parameters, including applicator type (3D-printed vaginal interstitial applicator [VIA] that accommodates a tandem and different combinations of ovoid sizes), were collected. The upper 1 and 2 cm VC volumes on pre-MRI were analyzed in relation to applicator type/size used. An XGBoost classifier was trained to distinguish VIA from non-VIA, incorporating imaging and clinical factors. To address class imbalance, we used SMOTE within a 10-fold cross-validation repeated 5 times. Model hyperparameters were optimized for the highest area under the ROC curve (AUC).
Results: Most patients had FIGO stage III/IV (75%) and squamous cell carcinoma (74%) disease. The same applicator type was used in all fractions for most patients (60%). VIAs were used in 26% of all fractions; of the remaining fractions, the combination of 2 cm ovoids bilaterally was used most frequently (39%), followed by dual mini (31%) and mini x 2 cm ovoids (19%). Applicator type was significantly associated with the upper 2 cm volume in each fraction (P < 0.01). In the 1st fraction, a trend was observed between the upper 2 cm volume and applicator type; median volumes were 4.2 cm³ (VIA), 6.7 cm³ (dual mini), 7.4 cm³ (mini x 2 cm), 9.1 cm³ (dual 2 cm), and 11.6 cm³ (2 x 2.5 cm). XGBoost achieved a mean cross-validation AUC of 0.69, with ~77% sensitivity for VIA identification. Feature importance analysis highlighted the upper 2 cm volume as the primary predictor, followed by the ratio of 2 cm to 1 cm volume.
Conclusion: Pre-MRI upper VC volumes correlate with brachytherapy applicator selection, with an upper 2 cm volume <5 cm³ benefiting from a smaller alternative applicator. While limited, our XGBoost model showed promise for predicting VIA use, potentially improving procedural efficiency and patient outcomes. Future work includes larger datasets and additional imaging metrics to refine predictive accuracy.