Ocular Disease Recognition

This project focuses on classifying ocular diseases from retinal fundus images using preprocessing techniques and deep learning models with transfer learning. It supports early diagnosis of conditions such as:

• Diabetic retinopathy
• Glaucoma
• Cataract
• Age-related macular degeneration (AMD)
• Hypertension-related changes
• Myopia
• Others

Dataset

We use the ODIR dataset, which includes:

• Multilabel classification (patients may have more than one disease)
• Metadata (age, sex, left/right eye images)
• Raw fundus images of varying quality The dataset and processed images are available via Google Drive:

• Training Images (Raw)
• Cropped Images
• Resized Images
• CLAHE Enhanced
• CLAHE + MIRNet Enhanced

Example images:

Preprocessing Step	Example Image
Raw
Cropped
Resized
CLAHE
CLAHE + MIRNet

Repository structure

File	Purpose
odir.csv	Metadata and labels for the fundus images
odir_resize.ipynb	Resizing pipeline to normalize image dimensions
odir_crop_images.ipynb	Crops noisy borders and focuses on retinal regions
odir_enhancement.ipynb	Enhances fundus image quality using histogram equalization, CLAHE, Gaussian filtering
hyp_resnet50_mirnet_msr.ipynb	Experiments with MIRNet enhancement + ResNet50 model
hyp_vgg16_mirnet_msr.ipynb	MIRNet + VGG16 pipeline
ODIR_balanced.ipynb	Class balancing using undersampling and training classifiers

Methods

Image Preprocessing

• Resizing and cropping for spatial consistency
• Enhancement with:
  • CLAHE (Contrast Limited Adaptive Histogram Equalization)
  • Gaussian blurring
  • Histogram normalization
• Deep enhancement via MIRNet for denoising and contrast recovery. MIRNet utilizes invertible building blocks, residual connections, attention mechanisms, and a perceptual loss function to produce visually pleasing images with improved brightness, contrast, and details while reducing noise and artifacts.

Classification Models

• ResNet50 and VGG16 with ImageNet weights
• Final classification head adapted for multilabel sigmoid outputs
• Training with binary cross-entropy loss and data augmentation
• Hyperparameter tuning using Keras Tuner (random search, hyperband, Bayesian optimization) with AUC as the primary metric

Evaluation Metrics

• Per-class Precision, Recall, F1-score
• AUC-ROC across all labels
• Confusion matrices and class activation maps (CAMs)

Results (Example)

Model	Preprocessing	F1 (Macro)	AUC-ROC
VGG16	MIRNet + CLAHE	0.71	0.91
ResNet50	MIRNet + CLAHE	0.74	0.93
MIRNet-enhanced images consistently outperformed raw image baselines.

How to Run

1. Clone the repo:

git clone https://github.com/YOUR_USERNAME/OcularDiseaseRecognition.git
cd OcularDiseaseRecognition

2. Ensure your environment has:

tensorflow, keras, opencv-python, numpy, pandas, matplotlib, seaborn, scikit-learn, keras-tuner

3. Run notebooks in order:
   • odir_resize.ipynb
   • odir_crop_images.ipynb
   • odir_enhancement.ipynb
   • Model experiments (hyp_*.ipynb)

Author

Sneha Dharne MS in Computer Science (ML) – Stevens Institute of Technology LinkedIn • GitHub

Additional Note

• The example images in the Dataset section are embedded using markdown images. They are hosted on GitHub (from the assets folder of the repo). The links in the table are direct links to the images.