Abstract
Severe stages of diabetes can eventually lead to an eye condition
called diabetic retinopathy. It is one of the leading causes of temporary visual disability and permanent blindness. There is no cure for this disease other than a proper treatment in the early stages. Five stages of diabetic retinopathy are discussed in this paper that need to be detected followed by a proper treatment. Transfer learning is used to detect the grades of diabetic retinopathy in eye fundus images, without training from scratch. The Kaggle EyePACS dataset is one of the largest datasets
available publicly for experimentation. In our work, an extensive study on the Kaggle EyePACS dataset is carried out using pre-trained models ResNet50 and DenseNet121. The Aptos dataset is also used in comparison with this dataset to examine the performance of the pre-trained models. Different experiments are performed to analyze the images from the different classes in the Kaggle EyePACS dataset. This dataset has significant challenges including image noise, imbalanced classes, and fault annotations. Our work highlights potential problems within the dataset
and the conflicts between the classes. A clustering technique is used to get informative images from the normal class to improve the model’s accuracy to 70%.
called diabetic retinopathy. It is one of the leading causes of temporary visual disability and permanent blindness. There is no cure for this disease other than a proper treatment in the early stages. Five stages of diabetic retinopathy are discussed in this paper that need to be detected followed by a proper treatment. Transfer learning is used to detect the grades of diabetic retinopathy in eye fundus images, without training from scratch. The Kaggle EyePACS dataset is one of the largest datasets
available publicly for experimentation. In our work, an extensive study on the Kaggle EyePACS dataset is carried out using pre-trained models ResNet50 and DenseNet121. The Aptos dataset is also used in comparison with this dataset to examine the performance of the pre-trained models. Different experiments are performed to analyze the images from the different classes in the Kaggle EyePACS dataset. This dataset has significant challenges including image noise, imbalanced classes, and fault annotations. Our work highlights potential problems within the dataset
and the conflicts between the classes. A clustering technique is used to get informative images from the normal class to improve the model’s accuracy to 70%.
| Original language | English |
|---|---|
| Publication status | Accepted/In press - 15 Nov 2022 |
| Event | The 3rd International Conference on Medical Imaging and Computer-Aided Diagnosis - University of Leicester, United Kingdom, Leicester, United Kingdom Duration: 20 Nov 2022 → 21 Nov 2022 Conference number: 3 https://www.micad.org/ |
Conference
| Conference | The 3rd International Conference on Medical Imaging and Computer-Aided Diagnosis |
|---|---|
| Country/Territory | United Kingdom |
| City | Leicester |
| Period | 20/11/22 → 21/11/22 |
| Internet address |
