RetinaRegNet

RetinaRegNet is a versatile registration model that achieves state-of-the-art performance in various retinal image registration problems. RetinaRegNet has the capacity to register images with large displacement deformation and minimal overlap. Unlike other deep learning-based registration methods, it does not require training on retinal images, making it a general approach applicable to other image registration problems. Our method is inspired by the paper titled "Emergent Correspondence from Image Diffusion", where the authors applied a latent stable diffusion model to generate robust Diffusion Features (DIFT) for finding semantic correspondences. We evaluated the performance of RetinaRegNet on three retinal image registration datasets: two public datasets, "FIRE" and "FLoRI21", and a privately owned dataset titled "LSFG".

This GitHub repository includes the latest implementation of our paper titled RetinaRegNet: A Zero-Shot Approach for Retinal Image Registration. Please note that this paper is an updated version of our previous paper titled RetinaRegNet: A Versatile Approach for Retinal Image Registration. Since both papers belong to the same study, we are using a single GitHub repository for both.

For additional queries about the model and its architecture, kindly refer to our paper titled "RetinaRegNet: A Zero-Shot Approach for Retinal Image Registration".

Requirements

The scripts were executed using a custom kernel configured with the list of packages mentioned below.

• torch==2.0.1
• torchvision==0.15.2
• jax==0.4.23
• jaxlib==0.4.23
• accelerate==0.23.0
• diffusers==0.20.2
• ipympl==0.9.3
• transformers==4.34.0
• xformers==0.0.22
• numpy
• scipy
• opencv-python
• matplotlib
• pyunpack
• Note: To install the aforementioned list of packages, kindly use the command pip install -r requirements.txt

Datasets

• Fundus Image Regsitration Dataset (FIRE)
• Fluorescein Angiography Longitudinal Retinal Image Registration 2021 Dataset (FLoRI21)
• Laser Speckle Flowgraphy Dataset (LSFG) -- To be released soon

Evaluation Steps

Evaluating our model on any of the mentioned datasets involves three simple steps, outlined below:

• Download a copy of the dataset using the provided link.
• Download the corresponding script for the dataset from this repository.
• Start executing the entire script with style.

Note:

• Before starting the model evaluation, ensure that the dataset and its corresponding script are saved in the same directory.
• The aforementioned scripts are customized for each dataset listed here, facilitating the automatic processing of all images and their corresponding ground truth landmarks.

References

Certain code cells were from this GitHub repository, as detailed in the paper titled "Emergent Correspondence from Image Diffusion".

Citations

Kindly cite our paper as follows if you use our code.

@article{SIVARAMAN2025109645,
title = {RetinaRegNet: A zero-shot approach for retinal image registration},
journal = {Computers in Biology and Medicine},
volume = {186},
pages = {109645},
year = {2025},
issn = {0010-4825},
doi = {https://doi.org/10.1016/j.compbiomed.2024.109645},
url = {https://www.sciencedirect.com/science/article/pii/S001048252401730X}
}

Contact

For any questions related to the code or its execution, please feel free to contact Vishal Balaji Sivaraman at [email protected].