Ocular Disease Recognition — Multi-Label Fundus Image Classifier

---

Overview

This project implements an end-to-end deep learning pipeline for automated detection of ocular diseases from retinal fundus images. The system performs multi-label classification across 8 diagnostic categories using the ODIR-5K dataset, and is designed with clinical applicability in mind — reporting sensitivity, specificity, and per-class AUC alongside standard accuracy metrics.

---

Disease Categories

Code	Disease
N	Normal fundus
D	Diabetes (diabetic retinopathy)
G	Glaucoma
C	Cataract
A	Age-related Macular Degeneration (AMD)
H	Hypertension
M	Myopia
O	Other pathologies

---

Dataset

ODIR-5K — Ocular Disease Intelligent Recognition
Source: Kaggle — andrewmvd/ocular-disease-recognition-odir5k

• 5,000 patients, 7,000 training images (left and right fundus per patient)
• Significant class imbalance: Normal and Diabetes ~2,100 cases vs Hypertension ~200
• Labels assigned at patient level; both eyes share the same label

---

Pipeline

Data loading & EDA
        ↓
Patient-level 5-fold split (prevents data leakage)
        ↓
TFRecord serialisation
        ↓
Preprocessing + augmentation (flip, brightness, contrast, saturation)
        ↓
Transfer Learning — EfficientNetV2B2 (frozen backbone)
        ↓
Fine-tuning — top 50% layers unfrozen
        ↓
Threshold optimisation on validation set (F1 and recall variants)
        ↓
Evaluation — AUC, F1, sensitivity, specificity, confusion matrices

---

Model Architecture

• Backbone: EfficientNetV2B2 (pretrained on ImageNet)
• Head: GlobalAveragePooling2D → Dense(256, ReLU) → Dropout(0.4) → Dense(8, sigmoid)
• Loss: Weighted Binary Cross-Entropy (pos_weight clipped to [1, 10])
• Activation: Sigmoid (multi-label)
• Input: 224×224 RGB, pixel range [0, 255] (EfficientNet internal rescaling)

Training strategy

Phase	LR	Loss	Monitor
Transfer Learning	1e-4	BinaryCrossentropy	val_loss
Fine-tuning	5e-5	Weighted BCE	val_recall

---

Results

Metric	Value
Test AUC (macro, multi-label)	0.8112
Test Accuracy	0.8778
Test Recall (sensitivity)	0.782
Test Precision	0.309
Macro F1	0.50
Micro F1	0.53

Per-class AUC

Class	AUC
Myopia	0.985
Cataract	0.950
AMD	0.897
Glaucoma	0.900
Hypertension	0.783
Normal	0.722
Diabetes	0.711
Other	0.568

---

Requirements

pip install tensorflow scikit-learn pandas numpy matplotlib seaborn pillow kaggle

---

Usage

1. Set up Kaggle credentials and download the dataset
2. Run cells sequentially in ocular_disease.ipynb
3. TFRecords are cached to avoid reprocessing on re-runs
4. Best models are saved as .keras checkpoints during training

---

Project Structure

├── ocular_disease.ipynb
├── data/
│   ├── dataset_remapped.csv
│   ├── labels.txt
│   └── plots/
│       ├── class_distribution.png
│       ├── confussion_matrices.png
│       ├── fine_tubed_plots.png
│       ├── sample.png
│       └── transfer_learning_plots.png
└── README.md

---

Course

Design of AI Systems
Authors: Josue Valenzuela Perez · Ludwig Alexandersson