Development of an artificial intelligence driven dose prediction pipeline for online adaptive magnetic...

The closed-off nature of most treatment planning systems (TPS) limits the potential for using artificial intelligence (AI) tools during online adaptive treatments. The aim of this study was to develop an AI-driven pipeline (AutoAdapt) for online planning of adaptive radiotherapy usable in a closed-off setting, providing optimal plan constraints derived from a population-based dose prediction model.

The AutoAdapt pipeline consists of a physics-aware Swin UNet transformer network for dose prediction trained on 266 magnetic resonance images from 25 prostate cancer patients treated with 60 Gy on a 1.5 T magnetic resonance linear accelerator. The predicted dose was used to calculate plan constraints that were subsequently fed into a commercial TPS. AutoAdapt was tested using ten unseen cases and compared to manual plans based on clinical objectives, time, and complexity.

While all plans were approved by a radiation oncologist, AutoAdapt met all clinical objectives in seven patients compared to ten when manually planned. AutoAdapt yielded a significantly lower D_0.035cm³ to the rectum (p = 0.01). Manual plans achieved a median rectum V_20Gy of 44% compared to 51% in AutoAdapt plans (p = 0.02). The pipeline only required a median of 29 s (7.5%) longer than the manual planners.

The developed pipeline resulted in high-quality plans, ready for clinical use without further adjustments. AutoAdapt prioritized maximum rectum dose over D_20% compared to manual planning, while requiring less manual work. In the future, AutoAdapt may be used to assist human planners and improve adaptive radiotherapy workflows.