GAS: A genetic atlas selection strategy in multi-atlas segmentation framework

Area of Science:

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

Background:

• Multi-Atlas based Segmentation (MAS) is effective but requires numerous atlases and precise registration.
• Atlas selection is critical for accurate segmentation, particularly for organs with significant anatomical and pathological variability.
• Current MAS methods face challenges in optimizing atlas subsets for diverse anatomical structures.

Purpose of the Study:

• To introduce a novel Genetic Atlas Selection (GAS) strategy for automated, optimal atlas subset selection in MAS.
• To enhance the accuracy and robustness of medical image segmentation, especially in cases of high organ variability.
• To improve the performance of MAS algorithms through intelligent atlas selection.

Main Methods:

• Proposed GAS strategy leverages image similarity and segmentation overlap for atlas selection.
• GAS identifies similar images to the target and employs a genetic algorithm to find near-optimal atlas subsets for each.
• Combined subsets from multiple similar images are used for the final segmentation of the target image.

Main Results:

• GAS demonstrated statistically significant performance improvements in MAS algorithms across single-label and multi-label segmentation tasks.
• Evaluated on prostate (whole and zonal) and left ventricle segmentation from MRI, GAS consistently enhanced accuracy.
• The method proved effective for both simple and complex segmentation scenarios involving anatomical variability.

Conclusions:

• The proposed Genetic Atlas Selection (GAS) strategy effectively automates and optimizes atlas selection for MAS.
• GAS enhances segmentation accuracy and reliability, particularly for challenging medical imaging datasets with high variability.
• This approach offers a significant advancement for improving the clinical utility of MAS techniques.