Artificial intelligence derived grading of mustard gas induced corneal injury and opacity

Abstract

Artificial intelligence (AI) has emerged as a transformative tool in ophthalmology for disease diagnosis and prognosis. However, use of AI for assessing corneal damage due to chemical injury in live rabbits remains lacking. This study aimed to develop an AI-derived clinical classification model for an objective grading of corneal injury and opacity levels in live rabbits following ocular exposure of sulfur mustard (SM). An automated method to grade corneal injury minimizes diagnostic errors and enhances translational application of preclinical research in better human eyecare. SM induced corneal injury and opacity from 401 in-house rabbit corneal images captured with a clinical stereomicroscope were used. Three independent subject matter specialists classified corneal images into four health grades: healthy, mild, moderate, and severe. Mask-RCNN was employed for precise corneal segmentation and extraction, followed by classification using baseline convolutional neural network and transfer learning algorithms, including VGG16, ResNet101, DenseNet121, InceptionV3, and ResNet50. The ResNet50-based model demonstrated the best performance, achieving 87% training accuracy, and 85% and 83% prediction accuracies on two independent test sets. This deep learning framework, combining Mask-RCNN with ResNet50 allows reliable and uniform grading of SM-induced corneal injury and opacity levels in affected eyes.

Keywords: Artificial intelligence; Cornea; Fibrosis; Pathology; Sulfur mustard.

Fig. 1

Workflow for the grading of SM-exposed corneas using CNN. The figure illustrates the end-to-end process of developing a deep learning model for grading corneal injury caused by SM exposure. The workflow consists of three main stages: (a) data acquisition and manual grading, (b) classification model development, and (c) SM-induced pathology grade prediction.

Fig. 2

Training and evaluation performance of the ResNet50 model on independent Test Sets 1 (a–c) and 2 (d–f). Panels (a, d) show the loss over epochs, with both training (black) and test set (red) losses decreasing and stabilizing, indicating effective model convergence. Panels (b, e) display the accuracy over epochs for training (black) and test sets (red), showing high and stable accuracy with minimal overfitting. Panels (c, f) present multiclass ROC curves for Test Set 1 and Test Set 2, with microaverage and macroaverage ROC-AUC values.

Fig. 3

Confusion matrices for six models (Baseline-CNN, VGG16, ResNet101, DenseNet121, InceptionV3, and ResNet50) evaluated on two test sets (Test Set 1 and Test Set 2) for classifying four severity classes: healthy, mild, moderate, and severe. Each cell in the matrix represents the number of true instances (rows) classified as predicted classes (columns). The diagonal values indicate the correct predictions for each class, whereas the off-diagonal values indicate misclassification. The shading represents the relative frequency of correct and incorrect classifications, with darker shades indicating higher counts.

Fig. 4

Representative examples of correctly classified corneal pathology grades by ResNet50 across independent test sets. The images in Test Set 1 and Test Set 2 demonstrate the model’s ability to accurately predict corneal health status across various pathology grades, including Healthy, Mild, Moderate, and Severe. For each example, the true label (True) and predicted label (Pred) are shown which illustrates robustness and generalizability of developed model across different datasets. ‘True’ refers to the expert-assigned grade based on manual annotation, while ‘Pred’ indicates the severity grade predicted by the AI model.

Fig. 5

Examples of misclassified corneal pathology grades by ResNet50 across independent test sets. The figure illustrates cases where the model predictions (Pred) did not align with the true labels (True) for different corneal pathology grades in Test Set 1 and Test Set 2.

Fig. 6

Demonstration of true and predicted corneal pathology grades for SM-exposed rabbit corneas. The first row shows the true pathology grades (healthy, mild, moderate, and severe), whereas the second row displays the grades predicted by the best-performing ResNet50 model. The third row presents corneal density heatmaps generated via Pentacam HR, validating the model’s predictions by illustrating corresponding changes in corneal density for each grade.