Fast Daily Auto-Contouring for Adaptive Cervical Cancer Treatments on the MR-Linac.
Abstract
Purpose
Effectiveness of EBRT depends highly on target coverage, which is challenged by inter- and intra-fraction anatomical motion. MR-guided radiotherapy (MRgRT) reduces the need for inter-fraction planning margins, by daily adaptive planning, for which daily recontouring remains a major bottleneck. Auto-contouring solutions can accelerate this process, by limiting manual delineation. This study evaluates three clinical-grade auto-contouring methods for their feasibility in MRgRT for gynecological cancers.
Methods
For ten gynecological patients treated on the Unity MR-Linac, the bladder, rectum and uterus were manually delineated on pre-treatment and daily MR-scans (6 fractions). Three clinical-grade auto-contouring solutions, were evaluated against rigid contour propagation: Monaco (v6.2.1.0) registration (Monaco-DIR): The current clinical solution employed in our department for cervical cancer daily contouring. Contour-guided EVolution (CG-EVO): In-house developed hybrid deep-learning/deformable registration algorithm. Monaco deep learning (Monaco-DL): Research version of Monaco with a DL-module for auto-segmentation. Combined/Hybrid: Best contour(s) between CG-EVO or Monaco-DL. Performance was assessed using DICE and 95%-Hausdorff (Hd95) distance relative to manual contours.
Results
All methods outperformed rigid propagation, in both DICE and Hd95. Monaco-DL showed the strongest bladder performance (DICEmedian=0.87). For the rectum and uterus, all methods achieved comparable DICE values (DICEmedian~0.8). Hd95median ranged between 4-9mm across structures and methods. Registration-based solutions (Monaco-DIR and CG-EVO) struggled with large-scale displacements/deformations. Monaco-DL was generally robust, but showed problems by atypical anatomies (e.g. catheters). Selecting the best contours between CG-EVO and Monaco-DL, resulted in DICEmedian>0.85 and Hd95median<5.9 mm across structures.
Conclusion
Auto-contouring methods improve the stability and accuracy of daily MR-guided adaptive radiotherapy for gynecological cancers. Depending on anatomical changes, either deep learning or deep learning-guided image registration performs best, and combining them can provide a robust, efficient workflow. Ultimately, evaluating residual need for manual edits and extending automation to tumor contours, will determine the true clinical impact of automated contour propagation within the adaptive workflows.