Hybrid Transfer Learning Model Combining Xception and NASNetMobile forDiabetic Retinopathy Detection

Abstract

Convolutional Neural Networks (CNNs) are sole of the various forms of deep
learning that have become effective tools in computer vision. Machine learning has
been transformed by Deep Neural Networks (DNNs) with many parameters; their
impact is particularly evident in network architecture. CNNs are very good at image
classification problems because of their ability to concentrate on objects in images
and extract information using spatial relationships. In this work, two robust
architectures Xception and NASNetMobile are integrated into a unique CNN model
for image categorization based on transfer learning. Using Xception and
NASNetMobile, the model uses photographs with specified dimensions, often known
as "best windowing of images," as input to categorize the images into two groups.
To avert overfitting issues in CNN, a dropout layer is introduced after the outputs of
these designs are concatenated using a concatenate layer. This research evaluates the
proposed model for an exceptionally challenging dataset associated with diabetic
retinal disease. The "Diabetic Retinopathy 224*224 Grayscale images" dataset,
which is part of the "APTOS 2019 Blindness Detection" dataset on Kaggle, has 3662
images, of which 1875 show abnormal cases and the residual 1805 show normal
instances. With an accuracy of 97.50%, precision of 96.39%, recall of 98.64%, and
F1-score of 97.36%, the model fared extremely well in this test.