An Open-source Protocol for Deep Learning-based Segmentation of Tubular Structures in 3D Fluorescence Microscopy...

Area of Science:

• * Biomedical image analysis
• * Computational biology
• * Deep learning applications

Background:

• * Segmenting tubular structures in 3D fluorescence microscopy images is crucial for understanding complex biological tissues.
• * Challenges include image complexity, variability, and quality issues, hindering accurate analysis.
• * Existing methods often require specialized programming skills, limiting accessibility for researchers.

Purpose of the Study:

• * To introduce an open-source, user-friendly toolbox for end-to-end segmentation of tubular structures in 3D images.
• * To provide researchers without formal programming training with advanced image analysis capabilities.
• * To improve the accuracy and efficiency of tubular network segmentation in biological tissues.

Main Methods:

• * Implementation of two deep learning architectures: 3D U-Net and 3D U-Net with attention mechanisms.
• * Development of a simulation-based data augmentation strategy to generate artificial microscopy images with realistic artifacts.
• * Systematic protocol guiding users through data augmentation, model training, evaluation, and inference.

Main Results:

• * Both 3D U-Net architectures demonstrated strong performance in segmenting tubular networks.
• * The attention U-Net slightly outperformed the standard U-Net when trained with augmented data.
• * The simulation-based data augmentation strategy significantly enhanced model performance, especially with minimal training data.

Conclusions:

• * The developed toolbox offers an accessible and efficient solution for segmenting tubular structures in 3D microscopy images.
• * The simulation-based data augmentation is a key innovation, enabling robust segmentation with limited data.
• * The toolbox democratizes advanced image analysis, empowering a broader range of researchers to study complex biological tissues.