Diabetic Retinopathy Detection

Automated diabetic retinopathy screening from retinal fundus images using CNNs and transfer learning.

Categories

Tech Used

Problem

Early DR detection is critical but manual grading is time-consuming and requires expertise; models must handle class imbalance, varied image quality, and device differences.

Approach

• Data preprocessing: resizing, normalization, contrast enhancement, and augmentation to improve robustness
• Modeling with CNN baselines and transfer learning (AlexNet, GoogLeNet, ResNet) with fine-tuning
• Training strategy: class-balancing, threshold tuning, and validation with standard classification metrics

Results

• Reliable DR classification performance with transfer learning
• Improved generalization via preprocessing + augmentation
• Reduced misclassification on challenging low-quality samples

Demo Videos