Retinal Abnormality Detector

Retinal Abnormality Detector
View on GitHub

Key Learnings

  • Understanding of AUC metrics and tradeoffs between precision and recall
  • Deep learning model architecture design for computer vision
  • Importance of balanced datasets in medical AI
  • The intense computational power required to train deep learning models

Project Overview

I hold this project in high regards since this was my first introduction into deep learning and computer vision. I was able to develop a deep learning model for classifying disease conditions from retinal scan medical imaging: - Implemented CNN architecture in TensorFlow - Explored novel AutoEncoder application for anomaly detection - Optimized model using AUC metrics - Conducted extensive hyperparameter tuning

Technical Details

- Developed multiple different architectures to try and get the model to train on my limited compute - The model wasn't effective and there are many improvements to make for the model now that I've had much more experience - I would use a pretrained model like ResNet50 and fine tune on this dataset - I would also use data augmentation to try and improve issues like class imbalance - I would also use cloud compute to efficiently train the model and allow for more experimentation

Technologies Used

Python TensorFlow OpenCV NumPy Computer Vision