4D Multi-atlas Label Fusion using Longitudinal Images

Area of Science:

• Medical Image Analysis
• Computational Neuroscience
• Radiology

Background:

• Longitudinal reproducibility in automated medical image segmentation is crucial but challenging due to manual tracing variability.
• Traditional 3D segmentation methods struggle with temporal variations inherent in longitudinal studies.
• Existing 4D longitudinal segmentation methods have limitations, such as application-specific energy functions or limited temporal smoothness considerations.

Purpose of the Study:

• To develop a novel 4D multi-atlas label fusion algorithm for general label fusion purposes.
• To enhance temporal consistency across all time points in longitudinal medical image segmentation.
• To improve the reproducibility of automated brain structure segmentation in longitudinal studies.

Main Methods:

• Proposed a novel longitudinal label fusion algorithm named 4D Joint Label Fusion (4DJLF).
• Incorporated temporal consistency modeling using non-local patch-intensity covariance models.
• Generalized a leading Joint Label Fusion (JLF) method to a 4D longitudinal framework, integrating spatial and temporal covariance across all time points.

Main Results:

• The 4DJLF algorithm demonstrated improved consistency in longitudinal segmentation compared to the original JLF approach.
• The method successfully modeled spatial-temporal consistency of atlases within a probabilistic framework.
• The proposed approach maintained sensitivity while enhancing consistency, validated using the same atlas set.

Conclusions:

• 4D Joint Label Fusion (4DJLF) offers a robust solution for improving longitudinal medical image segmentation reproducibility.
• The algorithm provides a general framework for label fusion that effectively models temporal dynamics.
• The open-source availability of 4DJLF facilitates its adoption and further development in the research community.