U-Net based vessel segmentation for murine brains with small micro-magnetic resonance imaging reference datasets

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Biology

Background:

• Accurate segmentation of murine vasculature is crucial for studying diseases like Alzheimer's and tumor progression.
• Manual or semi-automated segmentation is time-consuming and less accurate than deep learning approaches.
• Training deep learning models typically requires large labeled datasets, which are difficult to obtain in preclinical research.

Purpose of the Study:

• To develop and evaluate an open-source, shallow 3D U-Net architecture for automated segmentation of blood vessels in murine brains.
• To demonstrate the model's effectiveness with a small dataset and minimal labeled training data.
• To compare the model's performance against state-of-the-art vesselness filters.

Main Methods:

• Implementation of a shallow, three-dimensional U-Net architecture for vessel segmentation.
• Utilized a small dataset of 8 micro-magnetic resonance imaging (μMRI) stacks of mouse brains.
• Evaluated the model using cross-validation and two post-processing methodologies.

Main Results:

• The U-Net model achieved an average Dice score of 61.34% in its best configuration.
• The proposed methodology demonstrated faster and more reliable blood vessel detection compared to existing vesselness filters (average Dice score of 43.88%).
• The model requires only a small subset of labeled training data, significantly reducing annotation effort.

Conclusions:

• The developed shallow 3D U-Net provides an efficient and accurate solution for segmenting murine brain vasculature.
• This open-source tool facilitates preclinical research by automating a previously tedious task.
• The approach is suitable for studies involving tumor progression, angiogenesis, and vascular contributions to neurodegenerative diseases.