A Deep Learning-Based Segmentation Technique for Automatic Detection of Conjunctivitis from Conjunctival Eye Images

Abstract

Eye diseases are becoming common in day-to-day life and affecting all age groups of people. The ratio of ophthalmologists to patients suggests the need for an automated technique to detect eye diseases. Conjunctivitis is of many types, including adenoviral conjunctivitis, ocular drug toxic conjunctivitis, pollen-allergic conjunctivitis, bacterial conjunctivitis, and many others. Conjunctivitis can be automatically detected using conventional image processing techniques, but with lower accuracy and precision, and more computational time is required compared to deep learning and AI techniques. This paper presents a novel deep-learning-assisted segmentation technique for the automatic detection of conjunctivitis that overcomes the limitations of conventional methods. The proposed method uses Attention-U-Net++ with Transformer Encoder (Global Context), Swin / ViT CNN +, Transformer Monte-Carlo Dropout Layer Enabled at inference time, Uncertainty-aware and Segmentation Mask + Uncertainty Map, which provides better results with Accuracy= 90.3%, sensitivity= 0.87, specificity= 0.93, precision=0.88, recall=0.87, F1-Score= 0.89, ROC-Auc=0.96.