Similarity clustering-based atlas selection for pelvic CT image segmentation.

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Developing comprehensive medical imaging atlases is resource-intensive.
• Autosegmentation requires representative image datasets for training.
• Balancing anatomical diversity with resource limitations is crucial for atlas creation.

Purpose of the Study:

• To select a small, representative subset of pelvic CT images for an atlas.
• To enable autosegmentation of a target image set using a minimal atlas.
• To balance anatomical diversity with resource efficiency in atlas development.

Main Methods:

• Image preprocessing and registration were performed.
• Clustering algorithms selected a subset based on image intensity similarities.
• Robustness was tested via repeated clustering and random subset selection.

Main Results:

• A small, well-selected atlas set (5 images) provided coverage comparable to the full set (39 images).
• Selected atlases significantly outperformed random subsets in target set coverage (p < 1.0E-10).
• Retrospective atlas selection showed a marginal improvement over prospective selection.

Conclusions:

• A small, representative image subset can form an effective atlas for autosegmentation.
• This approach facilitates retrospective and prospective autosegmentation with reduced resource requirements.
• Utilizing multimodal imaging information in the atlas enhances structure definition.