Multi-atlas Segmentation as a Graph Labelling Problem: Application to Partially Annotated Atlas Data.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning for medical imaging

Background:

• Manual image annotation for multi-atlas segmentation is time-consuming and costly.
• Current methods cannot effectively utilize partially labeled atlases, limiting scalability.
• Reducing annotation effort is crucial for analyzing large medical image datasets.

Purpose of the Study:

• To develop a novel multi-atlas segmentation method capable of using partially labeled atlases.
• To formulate multi-atlas segmentation as a graph-labelling problem.
• To reduce the dependency on fully annotated atlases in medical image segmentation.

Main Methods:

• A graph-labelling approach using a Markov Random Field (MRF) formulation.
• Construction of a graph connecting atlases and the target image for label propagation.
• Extension of a continuous MRF optimization method to handle partially labeled data.

Main Results:

• The proposed method successfully performs multi-atlas segmentation with partially labeled atlases.
• Robust and accurate hippocampal segmentation was achieved on 202 subjects from the ADNI database.
• The method maintained accuracy even when only 20% of atlas slices were manually labeled.

Conclusions:

• The developed graph-labelling framework offers a unifying approach for multi-atlas segmentation.
• This method significantly reduces the burden of manual image annotation.
• The approach demonstrates high accuracy and robustness, particularly for large-scale neuroimaging studies.