Automatic Colorectal Cancer Screening Using Deep Learning in Spatial Light Interference Microscopy Data.

Area of Science:

• Digital pathology
• Biomedical optics
• Computational pathology

Background:

• Current surgical pathology relies on staining (e.g., H&E) for tissue visualization, introducing subjectivity and standardization challenges.
• Traditional methods hinder automated computational analysis due to variations in staining and imaging.
• Label-based staining requires external contrast agents, complicating data comparability across different settings.

Purpose of the Study:

• To introduce Spatial Light Interference Microscopy (SLIM) as a label-free imaging technique for pathology.
• To develop an automated colorectal cancer screening method using SLIM data and deep learning.
• To reduce human bias and enhance the comparability of imaging data in clinical settings.

Main Methods:

• Applied SLIM, a label-free microscopy technique measuring intrinsic tissue refractive index signatures.
• Utilized a mask R-CNN deep learning algorithm for automated analysis of SLIM images.
• Trained and validated the algorithm on a tissue microarray from 132 patients for colorectal cancer screening.

Main Results:

• Achieved 91% accuracy in gland detection.
• Reached 99.71% accuracy for gland-level classification (normal vs. cancerous).
• Obtained 97% accuracy in core-level classification of tissue specimens.

Conclusions:

• SLIM provides a label-free, objective imaging method for pathology.
• The combination of SLIM and deep learning enables accurate automated colorectal cancer screening.
• This integrated approach has the potential to become a valuable clinical tool for pathologists, enhancing speed and accuracy.