WE-E-213CD-06: A Locally Adaptive, Intensity-Based Label Fusion Method for Multi- Atlas Auto-Segmentation

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Computational Anatomy

Background:

• Atlas-based auto-segmentation (ABAS) is crucial for radiotherapy planning.
• Accurate contouring in ABAS relies on effective label fusion from multiple atlases.
• Existing methods like STAPLE have limitations in adaptive weighting.

Purpose of the Study:

• To introduce a novel locally adaptive, intensity-based label fusion approach for multi-atlas ABAS.
• To compare the performance of this new method against the STAPLE method.
• To enhance the accuracy and practical utility of ABAS in radiotherapy.

Main Methods:

• Developed a label fusion method using weighted averages of warped atlas label maps.
• Computed adaptive weights based on local correlation coefficients (LCC) at each image location.
• Incorporated neighboring atlas labels and selected the top k (25) with highest LCC to mitigate registration errors.
• Evaluated using ten H&N patient images with leave-one-out validation, comparing 7 structures.

Main Results:

• The proposed locally adaptive, intensity-based fusion method significantly outperformed the STAPLE method across all evaluated structures.
• Mean Dice value improvements ranged from 1.5% (right parotid) to 9% (right sub-mandibular gland).
• Demonstrated superior accuracy for structures including the mandible, parotids, sub-mandibular glands, brainstem, and spinal cord.

Conclusions:

• The novel locally adaptive, intensity-based label fusion method offers superior accuracy over STAPLE for ABAS.
• This improved accuracy enhances the overall performance and practical applicability of ABAS in radiotherapy.
• The findings suggest a significant advancement in automated contouring techniques for radiation oncology.