Prospective Application of an Artificial Intelligence Decision Support Tool for Applicator and Needle Selection in High-Dose-Rate Brachytherapy for Cervical Cancer Patients

Abstract

To assess the performance of an artificial intelligence (AI)-based decision support tool under differing thresholds to determine optimal applicator selection in a prospective clinical setting. Cervical cancer patients scheduled for high-dose-rate brachytherapy implants with intracavitary tandem and ring (IC) or interstitial tandem and ring (IC/IS) implants in a single tertiary cancer center were eligible for enrollment. Prior to the first brachytherapy implant, a diagnostic MRI was acquired. The clinical target volume and the expected inserted position of the intrauterine tandem and ring were identified and contoured on the T2-MRI. An in-house artificial intelligence-based predicted the need for an IC or IC/IS implant, based on target volume geometric features. For IC/IS implants, an optimal needle arrangement for target coverage was also predicted. Blinded to the AI outcome, a clinical determination was made by the clinician reviewing the MR image. AI algorithm prediction provided a confidence level associated with each decision. The algorithm performance for different confidence thresholds using the IC applicator of 50%, 60%, and 80% was investigated. Performance metrics of the initial clinical determination and the AI prediction were calculated based on the consensus optimal applicator determined from an assessment of planning dosimetry in the first fraction and clinical use for the final brachytherapy implants and fractions. The performance metrics were accuracy, precision, and recall. A total of 10 eligible patients were accrued between December 2020 and October 2022. Table 1 shows the initial clinical determination and consensus applicator AI predictions that were made with different confidence thresholds. The optimal confidence threshold (60%) yielded performance scores of 80%, 83.3%, and 83.3% for the accuracy, precision, and recall, respectively. The performance metrics were equivalent for the optimal confidence threshold and the initial clinical decision. The AI-based decision support tool shows strong predictive results for a clinical brachytherapy application that is important to patient outcome. This prospective study demonstrates that the algorithm's utility is a critical step in using AI-based tools clinically. Further work to determine the optimal brachytherapy applicator, based on treatment planning dosimetry, is required.