ML Imaging - Retinal Disease

Model Testing & Deployment Preview

Visual Explainability (Grad-CAM)

Grad-CAM visualizations used to highlight key image regions influencing model predictions.

Useful for model interpretability and trust-building with medical professionals.

Project Overview

Developed a deep learning-based diagnostic tool to aid early detection of ocular diseases; specifically glaucoma, diabetic retinopathy, and age-related macular degeneration; by classifying retinal fundus images through automated analysis.

The system applies transfer learning to large, labeled datasets, enabling accurate binary classification of retinal health and assisting clinicians with rapid, reliable diagnoses.

Project

TensorFlow, Keras, Python, Kaggle
Pre-trained CNNs (ResNet, Inception), transfer learning pipelines
Jupyter-based model evaluation and visualization

Tech Stack

Globally, millions suffer from vision loss due to late diagnosis of retinal diseases. Early detection is critical to prevent irreversible damage, yet many clinics lack the diagnostic tools or specialist access needed for timely intervention.

Why It Matters

We built a machine learning model to detect signs of ocular disease in retinal images. By leveraging transfer learning, the system identifies key patterns associated with glaucoma, diabetic retinopathy, and other conditions supporting early intervention, streamlined triage, and improved patient care.

Solution

Data Processing

Resized to 128×128, normalized pixel values, and built efficient pipelines; applied random flips, rotations, zooms, and brightness shifts during training.

Data Exploration

Retrieved retinal images via Kaggle API, cleaned dataset by removing duplicates, filtering corrupt/unlabeled files, and verifying annotation quality.

Class Imbalance

Only ~21% of images were healthy. Used oversampling, light augmentation, and class-weighted focal loss to address imbalance.

Model Development

Built a binary classifier using transfer learning with ResNet50. Initially froze the base, then fine-tuned all layers. Used Binary Focal Loss for imbalance, with early stopping and checkpointing.

Git Coming Soon