QD-RetNet: Efficient Retinal Disease Classification via Quantized Knowledge Distillation

Introduction

Retinal diseases can cause permanent vision loss if not diagnosed early, making efficient AI-based screening tools essential. However, most deep learning methods rely on heavy multi-modal inputs and large models, limiting their practicality. QD-RetNet addresses this by enabling lightweight, single-modality diagnosis through knowledge distillation and quantization-aware training.

This repository provides the official PyTorch implementation of our MIUA 2025 paper.

Figure: The proposed training pipeline consists of two steps. First, the fundus model teaches the quantized OCT model; second, the OCT model teaches the quantized fundus model. Only the quantized student model is used at inference, requiring just a single image modality (fundus or OCT).

---

Update

📌 [2025-05-12] QD-RetNet has been accepted for publication at the 29th UK Conference on Medical Image Understanding and Analysis (MIUA 2025 – Leeds).

---

Requirements

To set up the environment, run the following commands:

# Create a new conda environment with Python 3.10
conda create -n qd-retnet python==3.10

# Activate the environment
conda activate qd-retnet

# Install other dependencies
pip install -r requirements.txt

Implementation

DATA

Below are links to the original dataset repositories we used. Please follow their README instructions to download and organize the datasets:

• TOPCON-MM Dataset: xmed-lab/FDDM
• MMC-AMD Dataset: li-xirong/mmc-amd
• MultiEYE Dataset: xmed-lab/MultiEYE

This repository provides full setup and dataloaders for the TOPCON-MM dataset. For other datasets, you will need to modify the dataloader and folder structure accordingly.

---

Training Models

Train Single-Modal Teacher Models:

python train_oct.py
python train_fundus.py

Mutually train Quantized Student Models:

python main.py

(Optional) Train Full-Precision FDDM and ODDM Models for comparison:

python train_fddm.py
python train_oddm.py

Citations

If you find this repository helpful, please cite:

@inproceedings{kumar2025qd,
  title={QD-RetNet: Efficient Retinal Disease Classification via Quantized Knowledge Distillation},
  author={Kumar, Ashutosh and Verma, Manisha},
  booktitle={Annual Conference on Medical Image Understanding and Analysis},
  pages={279--292},
  year={2025},
  organization={Springer}
}

Acknowledgements

• We acknowledge the public availability of the datasets we used: FDDM, MMC-AMD, MultiEYE

• The knowledge distillation implementation is adapted from: https://github.com/xmed-lab/FDDM

• This research is supported by IIT (ISM) Dhanbad under the project FRS(220)/2024-2025/M&C.

Contact

For questions or collaborations, please reach out to:

• 📧 Ashutosh Kumar: kumarashutosh9694@gmail.com
• 📧 Manisha Verma: manisha@iitism.ac.in