Development of an AI Method for Rapid Immunohistochemical Stained Image Structure Auto-Segmentation
Abstract
Purpose
Manual segmentation of immunohistochemical (IHC) stained images is a time-consuming task that typically takes 1-2 workdays to segment all images needed for analysis. Deep learning-based methods were employed to create an AI model to automatically segment IHC stained images of the different subregions (crypts, stroma, and villi) of murine small intestine within seconds.
Methods
For each mouse, approximately 5 cm of jejunum was frozen in a “Swiss Roll” format, sectioned, stained, and imaged. Tissue sections were stained for gamma-H2AX (detecting double strand DNA breaks), total DNA content (using DRAQ5), and hypoxia content (using EF5 binding). A total of 61 EF5 stained images were manually segmented using ImageJ. 51 images were randomly shuffled into training (87.5%) and validation (12.5%) datasets, and 10 were evaluated by dice similarity coefficient (DSC). Multi-class masking was employed to assign a value and weight to each structure for segmentation predictions: background (weight = 0.0), crypts (weight = 4.0), villi (weight = 1.0), and stroma (weight = 4.0). The predicted segmentations can be manually modified before saving. The auto-segmented images were then used as binary mask overlays to segment their respective DRAQ5 and gamma-H2AX stained images using a separate in-house Python program.
Results
Overall, the model does well with segmenting the subregions of the intestine. The mean DSC for 10 images for each structure is reported: stroma 0.961, crypts 0.947, and villi 0.942. The segmentation (without modifications) takes seconds. However, modifications of varying degrees are generally required before the final segmentation is saved. Continued work is ongoing to further automate this process.
Conclusion
This algorithm has high potential to improve workflow for rapid IHC stained image structure auto-segmentation. Although this method was trained on intestinal IHC images, it is a general method that could be applied to other tissue types for segmentation as well.