Abstract
Purpose :
The corneal endothelium maintains stromal deturgescence through its barrier and fluid pump functions. The barrier function is conferred by its tight junctions (TJs) . The integrity of the TJs is frequently analyzed histochemically by immunostaining ZO-1. This study aims to grade the severity of damage to ZO-1 in the endothelium subjected to oxidative stress and cold storage based on machine learning.
Methods :
Immunocytochemical images of ZO-1 were drawn from studies on the endothelium subjected to cold storage (JOPT. 2022 ;38(10):664-681) and oxidative stress (JOPT. 2022 Nov 8.). The images were resized to 256x256 and segmented by deep learning (TVST, 2021, 1;10(13):27). Data augmentation was performed to increase the number of images. The images were then analyzed by Gabor filters for texture. Specifically, we generated a bank of 8 filters varying the theta parameter (0° to 157.5° in equal intervals) but other parameters were constant: kernel size = 7x7 pixels, Gaussian envelope (σ) = 0.5, wavelength(λ) = 5 pixels, spatial aspect ratio = 0.05 and Psi (Ψ) = 0. The mean and variance of the filtered images were computed as textural features. The textural features found to effective by ANOVA were subsequently employed to classify the images (i.e., grading the severity of damage) by decision tree, SVM, and KNN. We computed the performance metrics, including Precision, Recall, F1 scores, and accuracy.
Results :
An effective set of textural features computed from a bank of Gabor filters, as demonstrated by ANOVA (Fig. 1), enabled the classification of damage to ZO-1 into 4 classes: Control (1), Mild (2), Moderate (3), and Severe (4). Based on computed metrics for different classes, KNN outperformed with an average F1 score of 88% for all classes (92% - Severe, 89% - Moderate, 86% - Mild, 85% - Control). KNN also enabled the highest precision and recall. Table.1 gives a summary of the performance metrics for all the classifiers tested. Note an accuracy of 88% obtained with KNN declined significantly with SVM (68%) and decision trees (63%).
Conclusions :
The classification has become simple with features based on Gabor filters, yet it has achieved > 80% accuracy. Thus, we have developed a workflow that enables automatic stratification of damage to ZO-1. The approach can be applied to similar data during drug discovery or pathophysiological studies of other epithelia.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.