Abstract:
This research proposes a deep learning-based technique to classifying diabetic
retinopathy that employs three cutting-edge convolutional neural network models:
InceptionV3, Xception, and DenseNet201. The models were evaluated on a balanced
set of images that included photos classified into five categories: healthy, mild DR,
moderate DR, proliferate DR, and severe DR. DenseNet201 had the maximum
accuracy of 87%, having excellent precision and recall throughout all classes. With an
accuracy rate of 57%, InceptionV3 demonstrated reasonable performance, notably in
detecting reasonable and Proliferate DR. Xception outperformed InceptionV3 with a
60% accuracy rate, displaying greater precision and recall in most classes, particularly
Normal and Severe DR. The results show that DenseNet201 beats the other models,
making it a viable option for diabetic retinopathy identification. This study highlights
the potential of deep learning models to improve the accuracy and efficiency of
medical image processing, making them an important tool for early identification and
management of diabetic retinopathy. It is the most common cause of eyesight among
working-age adults globally. Early identification and therapy are critical for avoiding
significant vision loss. Traditional techniques of diagnosing DR include expert
ophthalmologists manually examining retinal pictures, which may be time-consuming
and subjective. With the introduction of deep learning techniques, especially
convolutional neural networks (CNNs), automatic image analysis has made great
progress. The thorough study covers a variety of investigations, from the examination
of varied datasets to the construction of real-time detection algorithms. However,
substantial shortcomings continue, including low dataset variety, a need for greater
model explainability, and difficulty in real-time implementation.
