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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Posterior penetrating eye injury poses risks of retinal tearing, detachment, and fibrosis, necessitating prompt visual assessment.
• Ophthalmoscopy, the current standard, has limitations in consistently detecting subtle retinal abnormalities post-injury.
• Automated image processing using machine learning (ML) offers potential for enhanced detection of retinal changes.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in identifying subtle retinal abnormalities following posterior penetrating eye injury.
• To compare the performance of different ML algorithms, including Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs), in detecting retinal injuries.

Main Methods:

• Ten rabbits underwent posterior penetrating eye injury and fundus photography.
• A dataset of 743 fundus images was randomly split into training (70%) and testing (30%) sets.
• Two ML model types (CNN and four SVM varieties) were trained and validated, with performance assessed using accuracy and ROC AUC.

Main Results:

• The Support Vector Machine (SVM) with a polynomial kernel achieved the highest accuracy (92.37%) and ROC AUC (0.96).
• The Convolutional Neural Network (CNN) model demonstrated strong performance with 91.03% accuracy and 0.91 ROC AUC.
• Other SVM models showed varied results, with linear kernel performing comparably and sigmoid/radial basis function kernels showing lower efficacy.

Conclusions:

• Trained ML algorithms can accurately identify retinal tears resulting from posterior eye trauma.
• This research represents a significant step towards developing computer-aided diagnostic systems for retinal injury detection.
• ML-based tools could improve the monitoring of eye injury and disease progression after penetrating trauma.