DIABETIC RETINOPATHY DETECTION USING DEEP LEARNING

Abstract

"Diabetic Retinopathy (DR) is a health condition affecting the human eye, particularly in
individuals with diabetes, causing damage to the retina and potentially leading to vision loss over
time. Currently, DR screening relies on manual assessment by ophthalmologists, a timeconsuming process. This project aims to address this issue by employing Deep Learning (DL), a
subset of Artificial Intelligence (AI), for the analysis of various DR stages.
The model utilized for this task is DenseNet, trained on a substantial dataset comprising
approximately 3662 training images. The objective is to automatically identify the DR stage,
classifying the images into detailed fundus images. The dataset used for training is sourced from
Kaggle (APTOS), comprising five stages of diabetic retinopathy labeled as 0, 1, 2, 3, and 4.
Patient fundus eye images serve as input parameters for the model.
The trained EfficientNetB0 architecture extracts feature from the fundus eye images, followed by
the softmax activation function producing the results.
This architecture achieves an impressive accuracy score of 0.9242 for DR detection. Finally, the
paper compares three Convolutional Neural Network (CNN) architectures, namely
EfficientNetB0, MobileNetV2 and DenseNet121.