Nonlinear microscopy and deep learning classification for mammary gland microenvironment studies.

Area of Science:

• Biomedical Optics
• Cancer Research
• Materials Science

Background:

• Tumor microenvironment and extracellular matrix (ECM) collagen morphology are critical in cancer progression.
• Label-free microscopy techniques like second harmonic generation (SHG) and polarization second harmonic (P-SHG) are valuable for studying ECM alterations.

Purpose of the Study:

• To investigate ECM deposition and collagen fibrillar orientation changes in mammary gland tumors using automated SHG and P-SHG microscopy.
• To develop and benchmark a deep learning model for classifying SHG images of mammary gland tissue.

Main Methods:

• Automated sample scanning with SHG and P-SHG microscopy.
• Development of two distinct image analysis approaches to differentiate collagen fibrillar orientation.
• Application of a supervised deep learning model, MobileNetV2 architecture with transfer learning, for image classification.

Main Results:

• Demonstrated two analysis methods to distinguish collagen fibrillar orientation changes in the ECM.
• Achieved 73% accuracy in classifying naïve versus tumor-bearing mammary gland SHG images using a fine-tuned deep learning model on a small dataset.

Conclusions:

• Automated SHG and P-SHG microscopy combined with deep learning offers a promising approach for analyzing ECM changes in cancer.
• The study highlights the potential of transfer learning with MobileNetV2 for classifying small biomedical image datasets.