3647 - Time-Efficient Structure Localization for Auto-Contouring in Radiotherapy

Purpose/Objective(s): Auto-contouring using artificial intelligence models is increasingly applied in radiation oncology to enhance the efficiency of contouring various structures, such as organs-at-risk and targets. Object detection (OD) is commonly used as a preprocessing step for auto-contouring models, providing initial localization of structures for precise delineation. However, conventional two-dimensional (2D) OD approaches are time-consuming, as they require inference across all axial slices of three-dimensional (3D) CT scans. Additionally, they lack a 3D anatomical understanding as they are trained on individual slices. While 3D OD models may provide better spatial representation, their slow and resource-intensive nature make them difficult to use in real-world settings. This study aims to develop an OD model framework that improves the time-efficiency of organ localization in CT images for auto-contouring models in radiotherapy.

Materials/Methods: To optimize processing speed while preserving anatomical information, CT images were preprocessed by projecting them onto axial, coronal, and sagittal planes. To enable visualization across different structures, multiple window settings such as lung, soft tissue, and bone were applied and stacked into RGB channels. YOLOv8 OD models, pre-trained on the ImageNet dataset, were individually trained for each plane, using 450 CT scans covering the head and neck, chest, abdomen, and pelvis.

Results: The model was validated on 88 CT scans using axial, coronal, and sagittal models. The proposed methods significantly improve inference speed, achieving an average speed of 4.2 ms per CT scan, compared to 214.1 ms for conventional 2D OD methods, making it 51 times faster. Additionally, it achieves comparable localization accuracy, with a 3D intersection over union (IoU) score of 0.848, compared to 0.830 for the 2D methods.

Conclusion: This study presents an OD framework that reduces inference time. By integrating efficient preprocessing strategies, the model provides a faster solution for enhancing organ localization in CT scans, supporting more accurate and efficient radiotherapy planning.