DR-detector

DR-detector detects diabetic retinopathy from color fundus images by combining transfer learning features with clinically interpretable retinal lesion features for automated classification.

Key Features:

• Integration of transfer learning and lesion features: Combines features from pretrained convolutional neural networks with clinically defined retinal lesion features to improve detection accuracy.
• Machine learning classifier: Uses Extreme Gradient Boosting (XGBoost) trained on combined CNN-derived and lesion features.
• Lesion segmentation: Segments and extracts exudates (EXs), microaneurysms (MAs), and hemorrhages (HEMs) using a UNET deep learning model.
• Transfer learning models: Employs pretrained VGG-16 and ResNet-50 to extract CNN features.
• Training dataset: Trained on 1,840 color fundus images from e-ophtha, retinal lesions, and APTOS 2019 Kaggle datasets with approximately half representing healthy eyes.
• External validation datasets: Validated on expert-annotated images from High-Resolution Fundus (HRF), MESSIDOR-2, and Indian Diabetic Retinopathy Image Dataset (IDRID).
• Performance: Reported accuracies include 100% with ResNet-50 combined with lesion features and 99.38% with VGG-16 combined features, and hemorrhage-only features achieving 99.38% accuracy.

Scientific Applications:

• Diabetic retinopathy detection: Automatic and accurate detection of diabetic retinopathy from color fundus images.
• Clinically interpretable prediction: Enables interpretation of predictions via explicit lesion features EXs, MAs, and HEMs.
• Cross-dataset validation: Applicable for screening and validation across datasets including e-ophtha, APTOS 2019, HRF, MESSIDOR-2, and IDRID.

Methodology:

Features extracted from pretrained VGG-16 and ResNet-50 and lesion segmentations produced by UNET (exudates, microaneurysms, hemorrhages) were combined and used to train an XGBoost classifier; training used 1,840 color fundus images from e-ophtha, retinal lesions, and APTOS 2019 and validation used HRF, MESSIDOR-2, and IDRID.