Rapid AI Auto-Planning Rivals Manual Expert Planning for Cervical Brachytherapy

Abstract

PURPOSE: The study aims to evaluate the quality and clinical acceptability of AI automated plans compared to manual clinical planning through blinded physician review, for cervical brachytherapy applicators. MATERIALS AND METHODS: Automated plans were generated using dose predictions from a U-Net with anatomical masks, dwell position location masks and applicator-specific 3D dose inputs (where dose was computed using uniform dwell times). Model data included 2005 brachytherapy plans from 7 implant types (train/validation/test split = 62/19/19%). Test set dose predictions were fed into an optimizer to produce automated plans. Randomized automated and clinical plan pairs were presented to 10 expert gynecologic brachytherapy physicians, who indicated plan preference, scored plans from 1-5 (where 5 indicates highest quality) and guessed which plan was automated. Five physicians from our center reviewed 130 plans in total across all 7 implants. Five external physicians from 3 other centers each reviewed 2 plan sets per implant type (70 plans). Auto-plan scores were compared between physician groups and compared to clinical plans with Wilcoxon signed rank tests (p<0.05 significant). RESULTS: Auto-plans were deemed better or equivalent in approximately 50% of cases for both physician groups, with highest preference rates for hybrid implants (>58% on average). Selection rates varied between physicians, often due to different prioritization of tumor coverage vs. organ sparing and/or loading preferences. Automated and clinical plans scored 4 (acceptable plan with clinically unimportant stylistic differences) on average (p>0.05 for all comparisons). Slightly reduced preference rates and scores for external physicians were attributed to stylistic planning differences not captured in model training data from our center. Physicians correctly identified about 50% of auto-plans, consistent with random chance, indicating indistinguishability from clinical plans. CONCLUSION: Our brachytherapy AI automated planning technology produced automated plans comparable in quality and indistinguishable from manual, clinical plans in a median of 1.4 minutes.

Ähnliche Arbeiten