Computer-Assisted Detection of Retinal Injury Following Ocular Trauma Using Machine Learning Algorithms

Posterior penetrating eye injury can lead to retinal tearing, detachment, and intraocular fibrosis that require timely intervention without compromising vision. Currently, ophthalmoscopy is the gold standard for visualizing the retina, even though one cannot consistently detect retinal abnormalities by relying on such a method alone. The objective of this work was to determine whether automated image processing systems based on machine learning (ML) could facilitate the identification of subtle changes in the retina following an injury.

Ten Dutch Belted rabbits were subjected to a posterior penetrating eye injury in the right eye via a 23-gauge needle and monitored using fundus photography. The dataset consisted of 743 full-color fundus images that were randomly split into training and testing groups at a ratio of 7:3. Two types of ML models, a convolution neural network (CNN) and 4 varieties of support vector machine (SVM), were constructed using training group images and validated with testing group images. The 5 models were compared using measures such as Accuracy, Precision, and the area under the receiver operating characteristic curve (ROC AUC).

The SVM polynomial kernel performed the best across all metrics, with 92.37% accuracy and ROC AUC of 0.96. The CNN model exhibited a relatively high accuracy of 91.03% and an ROC AUC of 0.91. The SVM linear kernel presented comparable but slightly lower metrics, with 83.86% accuracy and ROC AUC of 0.89. The lowest performing models were the SVM sigmoid (64.57% accuracy, ROC AUC=0.72) and radial basis function kernels (73.09% accuracy, ROC AUC=0.82).

Our study shows that trained ML algorithms can accurately identify retinal tears following trauma to the posterior segment of the eye. This is an important step toward developing computer-aided diagnostic tools that can be used to detect retinal injury and disease progression following posterior penetrating eye injury.