Automated olfactory bulb segmentation on high resolutional T2-weighted MRI

Area of Science:

• Neuroimaging
• Deep Learning
• Neuroanatomy

Background:

• Automated segmentation of the olfactory bulb (OB) is underdeveloped due to its challenging imaging properties.
• Advances in MRI resolution and deep learning offer new possibilities for accurate OB analysis.
• The OB plays a critical role in olfactory function, yet its automated assessment remains a challenge.

Purpose of the Study:

• To develop a novel, fast, and fully automated deep learning pipeline for accurate olfactory bulb segmentation.
• To evaluate the pipeline's performance on T2-weighted whole-brain MR images.
• To assess the pipeline's ability to generalize to independent datasets and replicate known biological effects.

Main Methods:

• A three-stage deep learning pipeline involving OB localization (FastSurferCNN) and segmentation (AttFastSurferCNN with self-attention).
• Training and testing on 620 manually annotated T2w images, with validation on the Rhineland Study and Human Connectome Project (HCP) datasets.
• Ensemble of predicted label maps for robust segmentation.

Main Results:

• High segmentation accuracy on the Rhineland Study (Dice: 0.852, VS: 0.910, AVD: 0.215 mm) and generalization to HCP data (Dice: 0.738-0.782).
• The pipeline sensitively replicated age-related OB volume effects (β=-0.232, p<.01).
• End-to-end processing time is under one minute per 3D volume on GPU.

Conclusions:

• The developed deep learning pipeline provides a validated, efficient, and scalable solution for automated olfactory bulb volume assessment.
• This method addresses the long-standing challenge of OB segmentation in neuroimaging.
• The pipeline's ability to generalize and detect biological effects highlights its potential for large-scale studies.