Geodesic shape regression based deep learning segmentation for assessing longitudinal hippocampal atrophy in dementia...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Deep Learning

Background:

• Longitudinal hippocampal atrophy is a key dementia marker.
• MRI segmentation errors limit precise atrophy quantification.
• Disease-unrelated longitudinal variability complicates analysis.

Purpose of the Study:

• To develop an accurate deep learning method for segmenting hippocampal morphology from longitudinal MRI.
• To mitigate longitudinal segmentation errors and enhance intra-individual consistency.
• To improve the measurement of hippocampal atrophy for dementia diagnosis.

Main Methods:

• Proposed GeoLongSeg, a diffeomorphic geodesic guided deep learning method.
• Integrated geodesic shape regression into a two-stage 3D U-Net segmentation network.
• Verified performance against state-of-the-art methods using test-retest reliability and variance ratio.

Main Results:

• GeoLongSeg demonstrated superior longitudinal morphological consistency compared to other pipelines.
• Segmentation results from GeoLongSeg showed significant discriminatory capability in distinguishing dementia patients from controls.
• Identified spatial and temporal local atrophy in the hippocampi of dementia patients.

Conclusions:

• GeoLongSeg offers an accurate and efficient approach for hippocampal segmentation from longitudinal MRI.
• The method enhances intra-individual morphological consistency, crucial for reliable atrophy measurement.
• Improved segmentation aids precise early-stage dementia diagnosis and patient stratification.