EfficientNet-DR: A Deep Learning Approach for Diabetic Retinopathy Detection and Classification

At a glance

Problem	Diabetic retinopathy is a top cause of preventable blindness, and many places lack eye specialists or costly equipment.
Idea	Use EfficientNet-B0, a small but accurate CNN, to grade eye scans from 0 (healthy) to 4 (severe).
Result	84.2% test accuracy across all 5 stages, with only 5.3M parameters and 20 ms per image.
Data	IDRiD, the first Indian-population retina dataset, with lesion-level labels
Tech	TensorFlow 2.8, Python 3.8, EfficientNet-B0 (ImageNet weights), Adam
Published in	2026 International Conference on Emerging Systems and Intelligent Computing (ESIC)
Where & when	February 2026 · Bhubaneswar, India
Publisher	IEEE (2026)
DOI	10.1109/ESIC68176.2026.11495823

The whole model is about 20 MB, so it fits on phones and edge devices
Runs with just 4 GB RAM and 2 GB storage
Blurry and dark scans are removed first; the rest are resized to 224×224 and contrast-normalised
Fewest parameters and fastest inference of the compared models
Next step: a portable retina camera with the model built in, for remote areas

Size and speed vs other published models

Model	Accuracy	Parameters	Inference
EfficientNet-B0 (this paper)	84.7%	5.3 M	20 ms
Hybrid-Net	88.0%	15.7 M	35 ms
EfficientNet (custom)	91.8%	11.5 M	28 ms
EDR-Net	92.8%	9.2 M	25 ms
U-Net + Inception	90%	10.3 M	40 ms

The five grades

Grade	Meaning
0	No retinopathy
1	Mild
2	Moderate
3	Severe
4	Proliferative (new abnormal vessels)

More numbers and details

Other models report higher accuracy on different datasets, but need 2-3× the parameters
Training: batch 32, Adam at 0.001, up to 100 epochs, early stopping after 10
Learning rate halves after 5 flat epochs (ReduceLROnPlateau, floor 1e-7)
Augmentation: 15° rotation, horizontal flip, 0.1 zoom, 0.2 brightness
Evaluated with accuracy, precision, recall, F1, AUC and a confusion matrix
EfficientNet scales depth, width and image size together, which keeps it small