Publication

Abstract : There is a growing concern in healthcare about the rise in eye disorders, necessitating innovative methods for early detection and precise classification. To improve the early detection of eye disorders, this work investigates the use of transformers to replace the traditional dense layers found at the top of pre-trained models. The proposed model utilizes convolution layers from pre-trained models to extract the activation attributes from fundus images and then fed to transformer block for final classification. To enhance the quality of the input data, a no-reference quality model called BRISQUE is applied to ensure that the model is trained exclusively on high-quality images, and contrast limited adaptive histogram equalization (CLAHE) is implemented to improve image contrast. The proposed model demonstrates robust performance, achieving a precision of 93.57%, recall of 93.71%, F1-score of 93.71%, and accuracy of 95%. This successful integration of convolutional neural networks with transformers showcases a promising approach for advancing the detection and classification of eye diseases.