Release v1.0: Enhancing Retinal Disease Detection with the Swin Transformer: A Comprehensive Comparative Analysis

This release contains the source code associated with the following academic study:

Title: Enhancing Retinal Disease Detection with the Swin Transformer: A Comprehensive Comparative Analysis. Authors: Ahmet Saygılı, Ömer Mutlu Atıcı
Published in: Concurrency and Computation: Practice and Experience (2025)
DOI: https://doi.org/10.1002/cpe.70366

Project Overview

This repository includes implementations and comparisons of five deep learning models for classifying retinal images:

• CNN (Convolutional Neural Network)
• CNN + Attention (CNN Enhanced with Attention Mechanism)
• CNN + Attention + LSTM (CNN Supported with Long Short-Term Memory Layer)
• Swin Transformer (Swin-based Transformer for Image Classification)
• Vision Transformer (ViT)

Key Evaluation Metrics

• Accuracy
• Precision, Recall, F1 Score
• GPU and RAM Usage
• Training Time
• Energy Consumption

Key Findings

• Swin Transformer: Highest accuracy (92%)
• Vision Transformer (ViT): Second-highest accuracy (88%)
• CNN + Attention + LSTM: Sustainable model with good performance (86%) and low resource utilization.

Dataset

Retinal image dataset used in this study is publicly available on Kaggle:

• [Medical Scan Classification Dataset (Retinal Imaging)](https://www.kaggle.com/datasets/arjunbasandrai/medical-scan-classification-dataset?select=Retinal+Imaging)

Dataset classes:

• Normal Eye (Healthy Retina)
• Glaucoma (Optic Nerve Damage)
• Diabetic Retinopathy (Diabetes-Related Eye Disease)
• Cataract (Clouding of the Eye Lens)

Image Preprocessing Steps

• Contrast Enhancement (Histogram equalization)
• Image Resizing:
  • CNN-based models: 256x256 pixels
  • Transformer-based models: 512x512 pixels
• Tensor Conversion
• Normalization

Requirements

To run these scripts, ensure Python and the following libraries are installed:

pip install torch torchvision timm scikit-learn matplotlib numpy opencv-python

Usage Instructions

Each deep learning model is provided in individual .py scripts:

• CNN model:

  python CNN.py
  

• CNN + Attention model:

  python CNN+Attention.py
  

• CNN + Attention + LSTM model:

  python CNN+Attention+LSTM.py
  

• Swin Transformer and Vision Transformer (ViT):

  python swin_vit.py
  

  (Specify the desired model within the script.)

Citation

If you use this code, please cite our paper as follows:

Saygılı, A. & Ömer, A. (2025). Enhancing Retinal Disease Detection with the Swin Transformer: A Comprehensive Comparative Analysis. Concurrency and Computation: Practice and Experience.

This project is licensed under the MIT License. Please see the LICENSE file for details.

Contact

For inquiries or assistance, please contact:

• [asaygili@nku.edu.tr, aticiomermutlu@gmail.com]

Full Changelog: https://github.com/asaygili/retina/commits/retina