Improved segmentation of collagen second harmonic generation images with a deep learning convolutional neural network

Area of Science:

• Biomedical Imaging
• Computational Biology
• Tissue Engineering

Background:

• Collagen fibers are crucial for tissue structure and function.
• Second Harmonic Generation (SHG) imaging visualizes collagen but faces segmentation challenges due to scattering and absorption at variable depths.
• Accurate collagen segmentation is vital for quantitative analysis in thick tissues.

Discussion:

• A U-Net convolutional neural network (CNN) was developed and trained for objective collagen-positive pixel segmentation in SHG z-stacks.
• The CNN's performance was rigorously benchmarked against conventional intensity-based thresholding techniques.
• The CNN demonstrated superior accuracy, especially at greater imaging depths where scattering and absorption effects are more pronounced.

Key Insights:

• The trained CNN effectively segments collagen fibers across a wide range of imaging depths in SHG microscopy.
• Deep learning-based segmentation significantly outperforms traditional intensity thresholding methods for SHG datasets.
• This method provides a robust solution for quantitative analysis of collagen in thick biological tissues.

Outlook:

• This CNN model can enhance quantitative SHG imaging applications in research and diagnostics.
• Future work may involve optimizing the CNN for other imaging modalities or tissue types.
• The approach holds potential for advancing studies in tissue regeneration, disease progression, and drug development.