Optical phenotyping using label-free microscopy and deep learning.

Area of Science:

• Biomedical imaging
• Computational pathology
• Cancer research

Background:

• Tissue phenotyping is crucial for understanding pancreatic cancer behavior and identifying therapeutic targets.
• Conventional phenotyping methods are destructive, time-consuming, and costly, limiting their practical application.
• There is a need for efficient, non-destructive techniques for tissue characterization.

Purpose of the Study:

• To develop an optical phenotyping approach for pancreatic cancer specimens.
• To combine label-free multiphoton microscopy with spatial transcriptomics and deep learning.
• To create a non-destructive method for classifying pancreatic cancer tissue subtypes.

Main Methods:

• Co-registered spatial transcriptomics, autofluorescence, and second harmonic generation microscopy data.
• Clustered tissue subregions into phenotypes based on transcriptomic signatures.
• Evaluated deep learning models for phenotype prediction using label-free imaging.

Main Results:

• A deep learning model achieved over 89% accuracy in classifying six tissue types from label-free microscopy images.
• The model demonstrated robust classification performance with area under the curve (AUC) values approaching 1.
• Successfully demonstrated the feasibility of optical phenotyping for pancreatic cancer specimens.

Conclusions:

• Optical phenotyping using label-free microscopy and deep learning is feasible for pancreatic cancer.
• This approach offers a non-destructive alternative to conventional tissue phenotyping methods.
• Future integration with gene expression data or advanced imaging techniques can enhance accuracy and clinical utility.