DeepSight: Retinal Disease Classification using EfficientNet-B3

Under Review at IEEE 2025

Overview

DeepSight is a deep learning framework for automated retinal disease classification from fundus images. The system combines EfficientNet-B3 as a feature backbone with a custom Attention Pooling mechanism that selectively focuses on diagnostically relevant retinal regions, improving robustness on imbalanced clinical datasets.

Architecture

Input Fundus Image (300x300)
        |
  EfficientNet-B3 Backbone (ImageNet pretrained)
        |
  Attention Pooling (spatial weighting over feature map)
        |
  Classifier: Linear -> SiLU -> Dropout -> Linear -> SiLU -> Dropout -> Output
        |
  Output: Normal / Diabetic Retinopathy / Other

Key Design Choices

• EfficientNet-B3 backbone — optimal accuracy/efficiency trade-off for medical imaging
• Attention Pooling — learns to focus on diagnostically relevant spatial regions instead of naive global average pooling
• Weighted Random Sampling — handles severe class imbalance in clinical retinal datasets
• Label Smoothing — reduces overconfidence on noisy medical labels
• Cosine Annealing LR — stable convergence over long training runs

Classes

Class	Description
Normal	Healthy retina
Diabetic Retinopathy	Diabetes-related retinal vessel damage
Other	Glaucoma, Cataract, Age-related Macular Degeneration

Dataset

ODIR-5K — Ocular Disease Intelligent Recognition dataset with 6,392 retinal fundus images across multiple disease categories.

Training

pip install -r requirements.txt
python src/train.py --data_dir data/retinal --epochs 40

Visualizations

Citation

@article{gobbilla2025deepsight,
  title={DeepSight: Retinal Disease Classification using EfficientNet-B3},
  author={Gobbilla, Jaya Krishna},
  journal={IEEE},
  year={2025},
  note={Under Review}
}