DigitAb: Domain-Adaptive Cell Type Prediction Method from Light Microscopy Images

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in medicine

Background:

• Histological stains like H&E are crucial for disease diagnosis and research.
• Immunostaining enhances cellular detail but is costly and complex.
• Multiplex imaging offers broad cellular coverage but faces accessibility challenges.

Purpose of the Study:

• To develop an accessible deep learning framework (DigitAb) for cell type classification directly from H&E stained kidney biopsy slides.
• To eliminate the need for specialized immunostaining or multiplex imaging assays in routine diagnostics.
• To enable scalable and cost-effective cellular segmentation for research and clinical pathology.

Main Methods:

• Trained a semantic segmentation model using DigitAb on ~3.5 million cells from 29 human kidney samples with Phenocycler-generated ground truths.
• Utilized adversarial domain adaptation to test DigitAb on unlabeled biopsy samples.
• Validated cell type predictions against the Banff schema for kidney transplant rejection and diabetic nephropathy characteristics.

Main Results:

• Achieved a balanced accuracy of 0.78 in classifying 10 cell types from H&E slides.
• DigitAb successfully predicted cell types from unlabeled kidney biopsy samples.
• Demonstrated high concordance with clinical gold standards for kidney transplant rejection and diabetic nephropathy.

Conclusions:

• DigitAb provides a scalable, accessible, and label-free solution for cellular segmentation using standard histology.
• The framework significantly reduces reliance on specialized assays, making advanced cellular analysis more widely available.
• This deep learning approach holds promise for improving kidney disease diagnosis and research.