Automatic Diabetic Retinopathy Classification with EfficientNet

Diabetic retinopathy is a leading cause of blindness in adults, and screening for it means an ophthalmologist reading fundus photographs one at a time — slow, and poorly reproducible between readers. Automating the grading is attractive for exactly that reason.

The hard version of the task is the five-class one: no DR, mild, moderate, severe, proliferative. Most published work at the time either collapsed it to two or three classes or topped out around 75% accuracy on five.

Approach

EfficientNet as the backbone, B4 and B5, pretrained on ImageNet and fine-tuned. The choice is about parameter efficiency — compound scaling gets high accuracy without the parameter count of the architectures earlier work had reached for.

Preprocessing matters more than it looks. The Kaggle dataset is a real screening archive, with wildly varying image quality. Contrast-limited adaptive histogram equalisation was tested on each colour channel separately:

CLAHE channel	Train	Validation
R	67.98%	63.60%
G	68.40%	64.65%
B	67.64%	64.70%
R+G+B	67.94%	63.90%
G + centre crop	71.50%	67.60%

Equalising only the green channel beats equalising all three — green carries most of the retinal vasculature contrast, and the other two mostly add noise. Cropping to the retina itself adds another few points.

Progressive resizing: train at 256 × 256 first, then continue at 512 × 512, 100 epochs total, with focal loss to cope with a dataset where 25,810 of ~35,000 images are class 0.

Results

Model	QWK	Weighted F1	Accuracy
Progressive EfficientNet-B4	0.7922	0.8269	83.87%
Progressive EfficientNet-B5	0.7931	0.8265	83.89%

Accuracy alone would be misleading on a dataset this imbalanced, so quadratic weighted kappa is the headline metric — it also has the right shape for the problem, since grading errors between adjacent severities matter less than errors across the whole scale.

Confusion matrices for EfficientNet-B4 and EfficientNet-B5 across the five diabetic retinopathy severity classes — Most mispredictions land one class away — which is the failure mode you want from a triage tool.

Two results worth keeping:

Progressive resizing helped, and helped the smaller model more — +3.47% accuracy on B4 against +1.07% on B5, compared with training at one image size.

B5 did not beat B4, despite 1.58× the parameters, which suggests B5 is simply overparameterised for this dataset. Bigger was not better here.

At 83.89% on five-class classification, this sits well above the ~75% that Pratt et al. and Lam et al. reported with larger architectures.