Data Diversity in Convolutional Neural Network Based Ensemble Model for Diabetic Retinopathy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Ophthalmology

Background:

• Diabetic retinopathy (DR) is a vision-threatening complication of diabetes requiring early detection.
• Automatic diagnosis systems (ADS) using biomedical imaging aid in identifying health problems.
• Ophthalmologists use fundus images (FI) to classify DR stages, crucial for preventing severe vision loss.

Purpose of the Study:

• To propose an ensemble model (EM) of three convolutional neural network (CNN) models for classifying diabetic retinopathy (DR).
• To address challenges of limited and imbalanced DR data by incorporating data diversity.
• To prioritize the early detection of Class 1 DR for timely disease control.

Main Methods:

• An ensemble model (EM) comprising three CNN models was developed.
• Data diversity was achieved through various augmentation and generation techniques, including affine transformation.
• The EM was trained and evaluated on limited and imbalanced DR datasets.

Main Results:

• The proposed CNN-based EM achieved a multi-class classification accuracy of 91.06%.
• The model demonstrated high performance metrics: 91.00% precision, 95.01% sensitivity, and 98.38% specificity.
• Results indicate superior performance compared to single models and existing methods.

Conclusions:

• The developed ensemble model effectively classifies diabetic retinopathy stages, particularly the critical early stage (Class 1).
• Data diversity strategies enhance the performance of CNN-based ensemble models for DR detection.
• This approach shows significant potential for improving automated diagnosis in ophthalmology and preserving patient vision.