Guided Context Gating: Learning to Leverage Salient Lesions in Retinal Fundus Images

An attention mechanism designed to suit the identification of micro and macro lesions in retinal fundus images for accurate diabetic retinopathy classification.

This paper has been accepted at IEEE International Conference on Image Processing (ICIP), 2024.

[IEEE Xplore][arXiv][Demo][Dataset]

Table of Contents

1. Motivation
2. Architecture
3. Quantitative Evaluation
4. Qualitative Evaluation
5. Usage
   • Setup
   • Inference on Custom Samples
   • Inference through Huggingface Spaces
6. Citation
7. License

Motivation

Diabetic Retinopathy (DR) grading is a highly challenging task due to the complexity of lesion variability and their contribution to different severity grades. Lesions such as Microaneurysms, Exudates, Hemorrhages, and Cotton Wool Spots vary significantly in size, shape, and appearance, often misleading the classification process. Some lesions are barely visible, while others are prominent, yet every minor lesion plays a critical role in determining the correct severity grade. Misclassifying even a single lesion can lead to incorrect grading, as the distinction between severity grades is often subtle and highly dependent on lesion characteristics.

Architecture

Our approach consists of three key sub-blocks:

1. Context Formulation: Captures the global spatial context of the image.
2. Channel Correlation: Identifies channel-wise relationships to enhance feature relevance.
3. Guided Gating: Uses the global context as a guiding signal to zero in on localized lesions.

Quantitative Evaluation

Class	Acc	Prec	Rec	F1	AUC	Support
No DR	95.65	91.67	95.65	93.62	99.43	23
Mild NPDR	100.0	100.0	100.0	100.0	100.0	01
Moderate NPDR	75.00	80.00	75.00	77.42	92.97	16
Severe NPDR	100.0	97.37	100.0	98.67	100	37
Very Severe NPDR	55.56	66.67	55.56	60.61	92.68	18
PDR	97.14	94.44	97.14	95.77	99.56	35
Advanced PDR	95.45	91.3	95.45	93.33	99.09	22
Macro	90.13	88.78	88.4	88.49	97.80	152
Weighted	-	89.51	90.13	89.73	98.64	-

Qualitative Evaluation

Usage

Setup

1. Creating a Virtual Environment:

   • The command python -m venv gcg_venv works the same on Windows, macOS, and Linux.

2. Activating the Virtual Environment:

   • On macOS/Linux, use source gcg_venv/bin/activate.
   • On Windows, use gcg_venv\Scripts\activate.

3. Installing the project:

pip install git+https://github.com/TejaCherukuri/Guided-Context-Gating.git

4. Install all the necessary dependencies:

   (Optional if using pip install git+<url>, as dependencies are installed automatically.)

pip install -r requirements.txt

Inference on Custom Samples

1. Prepare a list of test image paths (relative to your working directory).
2. Pass the list of paths to the predict() method from the inference pipeline.
3. The method will return predictions and generate visual outputs.

from gcg.pipelines import predict

# Example test image paths
image_paths = [
    'test_images/image_1.jpg',
    'test_images/image_2.jpg'
]

# Get predictions
predictions = predict(image_paths)
print(predictions)

Outputs

• Predictions: The predicted classes for the provided images.
• Heatmaps and Superimposed Images:
  • Generated for each test image.
  • Saved to the heatmaps directory in your working directory.

Example heatmap:

Inference through Huggingface Spaces

For those of you who prefer user interface over code, access the link to huggingface spaces [Demo]

Citation

If you find GCG useful for your research and applications, please cite using this BibTeX:

@article{cherukuri2024guided,
  title={Guided Context Gating: Learning to leverage salient lesions in retinal fundus images},
  author={Cherukuri, Teja Krishna and Shaik, Nagur Shareef and Ye, Dong Hye},
  journal={arXiv preprint arXiv:2406.13126},
  year={2024}
}

License

This project is licensed under the MIT License. See LICENSE for details.