ReVGG-R2Net

ReVGG-R2Net: Residual VGG with Recurrent Blocks for Medical Image Segmentation

A robust implementation of ReVGG-R2Net architecture combining VGG16 encoder with Residual Recurrent Convolutional blocks for medical image segmentation.

Dataset

Due to the large size of the dataset, we have hosted it on Google Drive.
You can access and download it using the following link:

👉 Download Dataset from Google Drive

Methodology

The models archtechture of our approach is illustrated below:

Architecture Overview

Encoder: VGG16 with ImageNet pretrained weights (all layers trainable) Core Innovation: Residual Recurrent Convolutional Blocks (R2 blocks) Decoder: Transpose convolution with skip connections Output: Sigmoid activation for binary segmentation

• Encoder: VGG16
• Core Innovation: Residual Recurrent Convolutional Blocks (R2 blocks)
• Decoder: Transpose convolution with skip connections
• Output: Sigmoid activation for binary segmentation

Key Features

• Residual connections with recurrent convolutions
• Batch size optimization during training
• Comprehensive evaluation metrics (Dice, Jaccard, F1-Score)
• Mac MPS GPU support
• Robust data preprocessing pipeline

Your dataset should have this format:

dataset/
├── Original
│   ├── image1.jpg
│   ├── image2.png
│   └── ...
└── Mask
    ├── image1.jpg
    ├── image2.png
    └── ...

Installation

!git clone https://github.com/FahimSultan-cyb/ReVGG-R2Net.git
import os, sys
root_path = os.path.join(os.getcwd(), "ReVGG-R2Net")
os.chdir(root_path)
!pip install -e .
!python setup_colab.py
!python download_pretrained.py

!pip install -r requirements.txt

Training

from scripts.train import train_model
model, history = train_model("path_of_your dataset")

Inference

from scripts.evaluate import evaluate_model

results = evaluate_model(
    model_path="pretrained",
    test_data_path="path_of_your dataset"
)

Evaluation

from scripts.evaluate import evaluate_model
results = evaluate_model(
    model_path="path/to/model.keras",
    test_data_path="path/to/test/data"
)

To run in google coolab (Less memory)

from scripts.train_colab import train_model_colab
model, history = train_model_colab(data_path="path/to/dataset")

Citation

If you find this code useful for your research, please consider citing:

@article{
    akter2026revgg,
    title={ReVGG-R2Net: Optimized recurrent framework for microscopic blood cell segmentation},
    author={Mst Shapna Akter and Md. Fahim Sultan and Tasmin Karim and Md. Shazzad Hossain Shaon},
    journal={Tissue and Cell},
    year={2026}
}

📌 Plain-text citation

Akter, M. S., Sultan, M. F., Karim, T., & Shaon, M. S. H. (2026).  
ReVGG-R2Net: Optimized recurrent framework for microscopic blood cell segmentation.  
Tissue and Cell, 98, 103189.  
https://doi.org/10.1016/j.tice.2025.103189