Integrated AI Contouring for Online Adaptation: Enabling Efficient Prostate Mrgrt and Beyond
Abstract
Purpose
Online adaptive prostate MR-guided radiotherapy (MRgRT) is time-sensitive, and contouring with structure preparation can require upwards of 15 minutes per fraction. While vendor-TPS provided contours can be useful, performance and consistency vary by site, protocol, and patient anatomy, and many programs benefit from specialized in-house models or alternative external AI solutions. Our goal was to develop a flexible integration pipeline that allows AI contours from internal or external sources to be incorporated directly into the Elekta Unity online adaptive workflow, enabling site and MD-tailored contouring while improving efficiency and maintaining clinician oversight.
Methods
We implemented an integration layer that accepts AI-generated structure sets from configurable sources and inserts them into the Unity online workflow at the point where vendor provided deformed structures are typically generated. The pipeline standardizes structure naming and performs automated mapping/validation against the preplan structures and delivers a ready-to-review contour set within the adaptive session. The system is designed to be model agnostic, supporting vendor, third-party, and in-house AI outputs with minimal workflow disruption. Clinicians retain full control through routine review and editing prior to plan approval.
Results
The integration reliably produced clinically usable contour sets within the online adaptive time window and reduced manual contouring workload during adaptive sessions by 32.2%. Across 300+ adaptive fractions, we evaluated the approach for operational efficiency gains and contour quality, demonstrating improved on-table workflow performance with consistent structures, clinically acceptable contour quality, and reduced editing burden.
Conclusion
A model-agnostic AI integration pipeline that interfaces with the Elekta’s MRLinac Unity adaptive workflow enables programs to leverage the most appropriate AI solution for their patient population and clinical standards, including specialized in-house models when vendor contours are insufficient. This framework improves efficiency, reduces contour-handling failure modes, and supports broader adoption and expansion of AI-assisted online adaptation for prostate MRgRT and beyond.