Stable Atlas-based Mapped Prior (STAMP) machine-learning segmentation for multicenter large-scale MRI data.

Area of Science:

• Medical Image Processing
• Machine Learning in Radiology
• Neuroimaging Analysis

Background:

• Machine learning (ML) segmentation is widely used in medical imaging.
• Variability in ML algorithm and intensity normalization choices impacts segmentation performance.
• Subcortical MRI segmentation accuracy and generalizability require further investigation.

Purpose of the Study:

• To evaluate the impact of ML algorithms and intensity normalization on subcortical MRI segmentation.
• To compare eight ML algorithms and eleven normalization strategies.
• To identify optimal configurations for robust and accurate brain MR segmentation.

Main Methods:

• Utilized a brain MR segmentation framework with eight ML algorithm configurations.
• Implemented eleven intensity normalization strategies, including Stable Atlas-based Mapped Prior (STAMP).
• Conducted experiments on down-sampled MR data and validated on a large multicenter dataset (n>3000).

Main Results:

• Ensemble-based ML algorithms (Random Forest, ANN) significantly improved segmentation accuracy.
• Random Forest demonstrated excellent agreement with expert manual delineations.
• STAMP normalization enhanced segmentation robustness across multicenter datasets, achieving high reliability.

Conclusions:

• The choice of ML algorithm and intensity normalization critically affects segmentation outcomes.
• ML-based segmentation tools are feasible for processing large-scale multicenter MR datasets with minimal expert intervention.
• The developed framework offers a reliable approach for automated subcortical MRI segmentation.