ESTRO 2025 - Abstract Book

S3116

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2025

Protocol-based one-click plan adaptations were performed for two cohorts: 15 esophagus and 30 head and neck (H&N) cases, each split 3:1 for soft-check threshold development/testing. Using nominal optimization objectives and fixed computation settings with deformed clinical contours on all available control-CTs, this approach produced 45 esophagus and 195 H&N adaptations. Each H&N case further received one clinically-prescribed off-line re-planning. Difference distributions for eight plan parameters were obtained from development data by comparing (I) clinical re plans with their nominal plan for H&N and (II) protocol-based adapted plans with their previous adaptation for both entities (Fig. 1A). Soft-check passing ranges were defined as the percentile intervals containing 95% of difference data for each parameter (PCI 95 ). Warnings/failures are triggered if differences exceed PCI 95 and 2×PCI 95 , respectively (Fig. 1B). Those entity-specific thresholds for protocol-based adaptations were tested on previously unseen plans, imposing pass rates ≥95% (PCI 95 benchmark) for all soft-checks. Results: Soft-check thresholds obtained from development data varied by adaptation technique (Fig. 2A) and target entity (Fig. 2B). Protocol-based adaptations, using the same field configuration and objectives as the nominal plan, showed greater similarity and smaller warning thresholds. Clinical re-planning, with more optimization flexibility and inter planner variability, yielded larger thresholds. The esophagus cohort showed larger parameter differences between adapted and nominal plans (higher warning thresholds) than the H&N cohort, likely due to more pronounced anatomical changes. Soft-check thresholds were validated on both cohorts’ testing data, with overall pass/warning/failure rates of 96.0%/2.9%/1.1% (H&N) and 98.2%/1.3%/0.5% (esophagus), surpassing the 95% passing benchmark (Fig. 2C). Certain parameters showed combined warning/failure rates >5%, attributed to smaller sample sizes in testing data.