Constructing datasets and training deep networks for the direct apoptotic cell identification from H&E staining images

Abstract Background Hematoxylin and eosin (H&E) staining is a simple and cost-effective method that can also be used to detect apoptotic cells, but even experienced pathologists often struggle to accurately identifying apoptotic cells using H&E tissue staining. This study was conducted to constructed a deep tunnel network capable of identifying and quantitatively analyze apoptotic cells in H&E-stained tissue sections. Methods Seventy-three samples from 4 cancer models across 3 animal species (number of patches = 44,189) were collected for analysis. A deep tunnel network was constructed to analyze apoptotic cells in H&E-stained tissue sections identifying and quantitatively using deep learning method. Results A total of 44,189 patches were used for network construction. The proposed network exhibited high accuracy (98.5%) in identifying apoptotic cells and was 17.4% more precise than the current gold standard. Moreover, compared to pathologists, the proposed network exhibited a 3-fold increase in accuracy and was 100 times faster in detecting apoptotic cells. Conclusions The proposed network enables accurate, efficient, and low-cost detection of apoptotic cells in H&E-stained sections, and can be readily integrated into existing pathology workflows to support reliable apoptosis assessment in primary care setting using standard H&E staining.