How many models/atlases are needed as priors for capturing anatomic population variations?

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Biomedical Imaging

Background:

• Medical image analysis often utilizes prior information from models/atlases to represent population variations.
• Key questions remain regarding the optimal number of atlases and subject images required for effective prior information encoding.

Purpose of the Study:

• To develop a method for determining the optimal number of models/atlases and subject images for medical image analysis.
• To address the under-researched questions of how many atlases and images are needed to optimally capture population variations.

Main Methods:

• A hierarchical agglomerative clustering algorithm was employed to partition images into groups based on dissimilarity.
• A Residual Dissimilarity (RD) measure was defined to evaluate partition quality.
• The variation of RD with the number of groups was analyzed to identify optimal partition breakpoints.

Main Results:

• A minimum of 5 to 8 groups (models/atlases) are essential for capturing diverse anatomic forms and body habitus.
• A minimum of 150 subject images are necessary to cover anatomical variations within a population for a specific body region.
• Body habitus variations were found to be a dominant factor in grouping, overriding factors like gender or moderate pathology.

Conclusions:

• The proposed method aids in constructing high-quality atlases from population image data.
• This approach is valuable for optimally selecting training image sets for deep learning strategies in medical imaging.
• Understanding the optimal number of atlases and images enhances the efficiency and accuracy of medical image analysis and model building.