Details

Autor(en) / Beteiligte

• Giridharan, Urja
• Janardhanan, Vedya
• Ravi, Prabha
• Hiremath, Basavaraj V

Titel

Classification of Fundus Images Using Modified Stacking Ensemble

Ist Teil von

• 2022 4th International Conference on Circuits, Control, Communication and Computing (I4C), 2022, p.330-335

Ort / Verlag

IEEE

Erscheinungsjahr

2022

Quelle

IEEE Electronic Library Online

Beschreibungen/Notizen

• Glaucoma and diabetic retinopathy are serious eye diseases that can impair vision if untreated. According to a WHO study, the number of diabetics who have retinopathy is expected to rise by more than 50% by 2030. The most frequent cause of permanent blindness is glaucoma, affecting at least 12 million people worldwide and leaving nearly 1.2 million people in India permanently blind. For early detection of glaucoma, fundus imaging datasets are collected from various sources and provided in open source for further research. The dataset is preprocessed through morphological closing and CLAHE, and data augmentation is considered to increase the size of the dataset. Fundus images are classified into three classes: glaucoma, diabetic retinopathy (DR), and normal through transfer learning algorithm of 5 CNN models: VGG19, ResNet50, MobileNetV3Large, EfficientNetB0, and VGG16. The accuracy is further improved through stacking ensemble procedure providing an accuracy of 96.7%. A novel technique of stacking ensemble is considered for the best 3 transfer learning models of ResNet50, VGG19, and MobileNetV3Large, with which the achieved accuracy is 98.3%.