Automatic cell counting from stimulated Raman imaging using deep learning.

Area of Science:

• Biomedical imaging
• Pathological analysis
• Deep learning applications

Background:

• Stimulated Raman scattering (SRS) microscopy enables label-free imaging of lipids and proteins in fresh tissues, crucial for tumor diagnosis.
• Accurate cell counting in SRS images is challenging due to low contrast and tissue heterogeneity.
• Existing methods lack efficiency for real-time analysis in clinical settings.

Purpose of the Study:

• To develop an automated cell counting framework for SRS images.
• To enhance tumor tissue characteristic analysis, cancer diagnosis, and surgical planning.
• To address the limitations of current cell counting techniques in complex biological samples.

Main Methods:

• A deep learning approach using a modified U-Net model for semantic segmentation.
• Integration of distance transform and watershed segmentation for cell instance segmentation.
• Validation on real human brain tumor specimens using SRS and H&E stained histological images.

Main Results:

• The framework achieved >98% Area Under the Curve (AUC) for cell counting accuracy.
• High correlation (R=0.97) was observed between SRS-based and H&E-based cell counts.
• The method demonstrated near real-time performance for automated cell counting.

Conclusions:

• The proposed deep learning framework offers a robust and accurate solution for automated cell counting in SRS images.
• This technique holds significant potential for improving diagnostic accuracy and surgical planning in oncology.
• The study highlights the efficacy of deep learning in advancing biomedical and pathological image analysis.