A Hybrid Attention-Based ML–DL Framework with Uncertainty Quantification for Automated Segmentation of Cotton Wool Spots and Major Conjunctiviti

Abstract

Human day to day activities and air pollution are the major cause for eye diseases which troubles from infant to old age persons. The ratio between the Ophthalmologists [3] and the number of patients suffering from eye diseases are increasing day by day, this leads technology to develop automatic technique for detection of eye diseases. In this paper a novel eye image segmentation technique has been discussed using AIML and deep learning algorithms, the proposed method is customized by deep learning U-Net++, CNN and Monte-Carlo Dropouts [1,2], and is built to segment CWS [4], lesions and enhances feature extraction in color fundus image. The architecture comprises of encoder and decoder with depth layer and various resolutions which are able to extract low and high level grained characteristics. A deep learning model is constructed for segmentation of conjunctiva nevus and major conjunctivitis by transformer Encoder (Swin/ViT) [6], Attention-U-Net++ Decoder using conjunctiva images or digital eye images.  The segmentation includes diseases like cotton wool spots, conjunctiva nevus and major conjunctivitis using conjunctiva images and fundus images. Data augmentation is considered to ensure reliable results and to avoid over fitting of the model by increasing generalization capability. The segmentation model achieved a accuracy of about 96.32% and a dice coefficient of 88.60% in lesions segmentation for fundus images. The segmentation model for conjunctiva images provide test accuracy of 82.40% with a specificity of 94.55%.