Automated Multi-Class Classification of Anterior Segment Eye Diseases Using Deep Learning

Abstract

Accurate and timely detection of anterior segment eye diseases is critical for preventing permanent loss of sight. In this paper, we present a framework for applying deep learning to identify five different types of external eye-related diseases at once, namely: cataracts; corneal ulcers; pterygia; conjunctivitis; and eyelid abnormalities. Labeled clinical images of the eyeball were collected from both publicly available and clinician-curated sources. We tested three different deep Learning-Based Convolutional Neural Network (CNN) models: (1) a custom designed 5-layer CNN, (2) VGG16 implemented with transfer learning, and (3) EfficientNet-B0. All three models were tested under the same preprocessing algorithms, stratified data splitting, and augmentation techniques. The multi-class performance of each model was measured using the four common performance metrics (i.e., accuracy, precision, recall, and F1-score). The experimental results show that the best performing model was the EfficientNet-B0, which achieved a maximum accuracy of 90.5% with also demonstrating superior computational performance compared to the other two CNNs. It is suggested that the proposed framework could be used for automated diagnostic screening and/or tele-ophthalmologic applications.