Aldahmashi/DR-EfficientNetB0

Diabetic Retinopathy Grading — EfficientNetB0

Fine-tuned EfficientNetB0 for 5-class diabetic retinopathy severity grading from retinal fundus photographs, trained on the APTOS 2019 Blindness Detection dataset.

---

Model Details

Property	Value
Base model	EfficientNetB0 (ImageNet pretrained)
Framework	Keras 3.13.2
Input shape	224 × 224 × 3 (RGB)
Output	5-class softmax
Parameters	~5.3M
File size	33.4 MB (float32)

---

DR Severity Classes

Grade	Label	Clinical Meaning
0	No DR	No signs of diabetic retinopathy
1	Mild	Microaneurysms only
2	Moderate	More than microaneurysms, less than severe
3	Severe	Extensive hemorrhages, venous beading
4	Proliferative	Neovascularization or vitreous hemorrhage

---

Training Setup

Dataset: APTOS 2019 — 3,662 fundus images across 5 severity grades

Two-phase fine-tuning:

• Phase 1 (epochs 0–9): EfficientNetB0 backbone frozen, classification head trained from scratch
• Phase 2 (epochs 10–21): Full network unfrozen and fine-tuned with reduced learning rate

Preprocessing pipeline: Rescaling (÷255) → per-channel Normalization

Regularization: Dropout(0.3) before the final Dense layer

---

Evaluation Results

Evaluated on 550 held-out validation images from APTOS 2019.

Overall accuracy: 72% · Macro F1: 0.57 · Weighted F1: 0.73

Class	Precision	Recall	F1-score	Support
No DR	0.95	0.94	0.95	271
Mild	0.33	0.62	0.43	56
Moderate	0.74	0.47	0.57	150
Severe	0.34	0.55	0.42	29
Proliferative	0.55	0.41	0.47	44

Note: The APTOS 2019 dataset is heavily skewed toward No DR (49% of samples). The model performs strongly on the dominant class (F1 = 0.95) but struggles on minority classes — a known limitation of unimodal image-only approaches on imbalanced medical datasets. This serves as the baseline motivation for multimodal DR grading research incorporating metabolic context (HbA1c, diabetes duration).

---

How to Use

import keras
import numpy as np
from PIL import Image

# Load model
model = keras.saving.load_model("final_model.keras")

LABELS = ["No DR", "Mild", "Moderate", "Severe", "Proliferative"]

def predict(image_path: str):
    img = Image.open(image_path).convert("RGB").resize((224, 224))
    arr = np.expand_dims(np.array(img, dtype=np.float32), axis=0)
    probs = model.predict(arr)[0]
    label = LABELS[np.argmax(probs)]
    confidence = float(np.max(probs))
    return label, confidence

label, conf = predict("fundus.jpg")
print(f"Prediction: {label} ({conf:.1%})")

---

Limitations

• Class imbalance: Performance degrades on minority classes (Severe, Proliferative) due to dataset skew
• Boundary confusion: Adjacent severity grades (Mild ↔ Moderate, Severe ↔ Proliferative) are frequently confused — a property inherent to unimodal retinal image analysis
• Unimodal: Does not incorporate metabolic or clinical context that clinicians rely on
• Dataset scope: Trained solely on APTOS 2019 — generalization to other fundus camera types or populations is untested
• Not for clinical use: This model has not been validated for medical diagnosis

---

Citation

@misc{aptos2019,
  title   = {APTOS 2019 Blindness Detection},
  author  = {Asia Pacific Tele-Ophthalmology Society},
  year    = {2019},
  publisher = {Kaggle},
  url     = {https://www.kaggle.com/c/aptos2019-blindness-detection}
}

---

Author

Nasser Aldahmashi
AI specialization student · Asia Pacific University (APU), Kuala Lumpur
🤗 huggingface.co/Aldahmashi