TinyOcuNet

A Lightweight Deep Learning Model for Ocular Disease Classification

Model Design

Classification Type	Architecture / Features	Purpose
Binary Classification	Derived from VGG, with fewer convolutional layers for efficiency	Classify Cataract vs. Normal Eye
Multi-Class Classification	Incorporates Squeeze-and-Excitation (SE) blocks at the end of each convolutional block; learns discriminative channel-level features	Generalizes to multiple ocular diseases

Training uses mixed precision to reduce model size and training time while maintaining strong performance.

---

Usage

1. Install Dependencies

pip install -r requirements.txt

2. Run the Streamlit Demo

streamlit run ./demo/main.py

---

Loss and Accuracy Curves

Results for Multiple Ocular Eye Disease Classification.

Best Epoch	Train Loss	Test Loss	Train Acc	Test Acc
91	0.1931	0.1432	92.63	95.21

Training Curves:

Loss Curve	Accuracy Curve

Confusion Matrix	ROC Curve

Pretrained Models

Pre-trained weights are available in the models/ directory. To load the trained model (in Colab or locally):

model.load_state_dict(torch.load("models/MultipleEyeDiseaseDetectModel.pth"))

---

If you find this project useful, consider starring the repository to support further development.