Hierarchical model-based object localization for auto-contouring in head and neck radiation therapy planning

Area of Science:

• Medical Imaging
• Radiation Oncology
• Computer Vision

Background:

• Organ at risk (OAR) segmentation is crucial for radiation therapy (RT) treatment planning.
• Automatic anatomy recognition (AAR) offers a two-step approach: object recognition and delineation.
• This research focuses specifically on improving the recognition step within AAR.

Purpose of the Study:

• To introduce novel features for the AAR-recognition (AAR-R) approach.
• To integrate texture and intensity information for enhanced object recognition.
• To utilize an optimal spanning tree for hierarchical recognition and minimize errors.

Main Methods:

• Developed AAR-R by combining texture (maximum probability of occurrence) and intensity features.
• Employed an optimal spanning tree to establish a recognition hierarchy.
• Tested the approach on large-scale computed tomography (CT) datasets from head and neck cancer patients.

Main Results:

• Texture and intensity integration significantly reduced location errors for gland-related OARs.
• The novel approach achieved a location error under 4 mm for good quality images.
• Recognition performance was evaluated considering image quality, a new concept in this context.

Conclusions:

• The enhanced AAR-R method improves the accuracy of OAR recognition in RT planning.
• Combining texture and intensity information is effective for precise object localization.
• The approach shows promise for accurate OAR recognition in both serial and non-serial CT studies.