3589 - Clinical Development and Implementation of Zero-Click AI-Based Auto-Segmentation Workflows for Pediatric Radiation Therapy

Purpose/Objective(s):

To develop a zero-click approach of having custom-made AI-based workflows automatically segment simulation CT images in accordance with the AAPM TG-263 nomenclature in preparation for treatment planning.

Materials/Methods:

AAPM TG-263 compliant region-of-interest (ROI) templates with unfilled target, organs-at-risk (OAR), alignment, and override contour structures were created for different anatomical sites which are used to achieve consistent contour color, type, and contour codes. A commercial AI contouring tool utilizing four base models (head and neck, thorax, abdomen, and pelvis) were combined with the ROI templates to generate workflows capable of auto-segmentation for brain, head and neck, thorax, abdomen, and pelvis female/male anatomical sites. A hybrid craniospinal irradiation (CSI) workflow was created by running the AI contouring tool’s head and neck and thorax base models with post-processing to create an intermediate step RT structure file with a unique series description. This series description automatically triggers atlas-based auto-segmentation, followed by post-processing as well as loading of the ROI template to create an RT structure file that can be edited further if needed prior to CSI planning. Atlas-based auto-segmentation was used to include an optimization target volume which encompasses vertebral bodies (OTV_VB) that need to receive uniform dose levels in young children and a “CTV_Spine” structure which includes the spinal canal with lateral nerve roots. A simple extremity workflow was also created that applies to the ROI template and segments the skin and bone contours. A filter list was developed to check the series description, modality, and patient sex DICOM parameters of CT images exported from CT simulators to determine the appropriate workflow to automatically run for each CT scan.

Results:

A zero-click AI-based solution for contour auto-segmentation greatly reduced the time spent contouring from 2-4 hours to 20-60 minutes with most OARs being clinically usable with minimal changes despite the AI contouring tool being trained mostly on adult patients. This solution has been in clinical use since 2024 for over 100 pediatric patients and contour accuracy improvements were also observed for the hybrid CSI workflow compared to the existing atlas-based auto-segmentation method for OARs such as the kidneys which usually required significant edits for some patients for which the atlas database had few if any matching patients.

Conclusion:

Implementing such a zero-click solution in a radiation oncology clinic can significantly save time and effort with minimal modification by clinical staff for the contouring process in radiation treatment planning. This work also shows that hybridization of auto-segmentation techniques is possible and has the potential to increase segmentation accuracy and save time.