International journal of radiation oncology, biology, physics | 2021
Feasibility of Knowledge-Based Dose Predictions to Inform Needle Supplementation in Cervical Brachytherapy.
Abstract
PURPOSE/OBJECTIVE(S)\nSupplementing intracavitary (IC) applicators with interstitial needles enables more customized dose distributions and allows treatment of more complex cervical cancer cases. However, there are currently no tools to assist physicians in deciding when needles are needed. The purpose of this study is to test the feasibility of using knowledge-based dose predictions to guide needle supplementation.\n\n\nMATERIALS/METHODS\nKnowledge-based models for tandem-and-ring/ovoids (T&R/T&O) applicators were produced to predict OAR dose with 0.4-0.6Gy precision. Models were applied to 32/36 T&O/R patients with 0-3 added needles (126/145 treatment fractions). Prediction accuracy was verified by replanning all cases as IC-only, guided by OAR predictions, to meet HRCTV D90 85-90Gy, and if possible, EQD2 D2cc aims: bladder < 80Gy, rectum < 65Gy, sigmoid < 70Gy. A case was classified as hybrid (needles needed) if any replanned OAR D2cc > aim; if all aims were met it was classified as IC. We generated a receiver operating characteristic (ROC) curve and used the Youden index to determine the optimal classification threshold. To explore how clinical implementation could proceed, we used fraction 1 predictions to guide needle choice for fraction 2, fraction 2 to guide fraction 3, etc., with choice for fraction 1 based on physician. We applied this workflow to our dataset and compared model results to physicians applicator choice, with replan OAR results determining the true classification.\n\n\nRESULTS\nThe area under ROC curve for T&O/R model was 0.79/0.81, proving high classification accuracy when comparing predictions to same-day replans. Optimal thresholds were 113%/105% of dose aims for T&O/R. When predicting needles for subsequent days, 61/60% of T&O/R cases were predicted correctly, compared to 57/63% accuracy of the physicians choice. Worse prediction accuracy for T&R was likely due to daily variability in bladder fill and contouring. Accuracy of IC prediction (78/69%) was better than hybrid (53/49%); this is consistent with trends observed in ROC. There likely is a band of uncertainty around the dose aims (e.g., ± 5% of aims) where physicians would accept either choice depending on preference or other factors. E.g., a physician may prefer to avoid needles if doses are very close to aims for patient comfort. In 25% of cases, replan D2cc fell within this band, making accurate binary classification more challenging and perhaps less crucial in these cases. Model thresholds could be tuned to better reflect physician preferences.\n\n\nCONCLUSION\nThe suggested workflow using model-predicted dose to guide needle choice proved feasible, with accuracy comparable to experienced physicians. Predictions for same-day needle usage were more accurate than for subsequent fractions due to anatomical and contouring variability. Dose predictions could standardize optimal applicator selection in gynecologic brachytherapy, which is traditionally based on physician experience.