We propose MaDoUNet, a lightweight dual‑encoder architecture that combines EfficientNetB4 and DenseNet121 to extract complementary high‑level semantic and low‑level structural features. EfficientNetB4 is enhanced with depthwise convolutional transformers to improve global context modeling, while DenseNet121 is integrated with a novel VSS‑Mamba module for effective temporal‑spatial feature representation. A refined decoder and a composite loss function—combining Binary Cross‑Entropy and Dice Loss—further enhance segmentation performance and address class imbalance.
Experimental results demonstrate that MaDoUNet achieves consistent improvements across key metrics, with average gains of approximately 3–5 % in Dice score, IoU, and precision compared to existing methods. With its strong accuracy and computational efficiency, MaDoUNet offers promising potential for real‑time clinical applications in polyp segmentation.
The proposed MaDoUNet employs a dual-encoder structure combining EfficientNetB4 and DenseNet121. It incorporates depthwise transformer blocks and a custom VSS-Mamba module to enhance both global and local feature representation.
MaDoUNet demonstrates significant improvements in segmentation performance over existing models. It consistently achieves higher Dice scores, IoU, and Precision metrics across test datasets such as Kvasir-SEG and CVC-ClinicDB.
Below is a qualitative result showcasing input image, ground truth, and predicted mask side-by-side:
Quantitative Evaluation:
| Dataset | mIoU | DSC | Recall | Precision | F2 Score |
|---|---|---|---|---|---|
| Kvasir-SEG | 0.8536 | 0.921 | 0.9403 | 0.9025 | 0.9310 |
| CVC-ClinicDB | 0.9027 | 0.948 | 0.9556 | 0.9422 | 0.9520 |
📌 Note: These results were obtained on the Kvasir-SEG and CVC-ClinicDB datasets using a 80/20 train-validation split.
Follow the steps below to train, test, visualize, or run inference using MaDoUNet.
We recommend using Python 3.10 or later.
pip install -r requirements.txtOptionally, install CPU-only PyTorch:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpuPrepare your dataset (e.g., Kvasir-SEG) in the following format:
dataset_root/
├── images/
│ ├── img1.jpg
│ ├── img2.jpg
│ └── ...
└── masks/
├── img1.jpg
├── img2.jpg
└── ...
python main.py --train \
--image_dir /path/to/images \
--mask_dir /path/to/masks \
--run_instance_name your_run_nameAll training results (model checkpoint, logs) will be saved under:
results/train_results/your_run_name/
python main.py --visualize \
--checkpoint_path results/train_results/your_run_name/madounet_best.pth \
--image_dir /path/to/images \
--mask_dir /path/to/masks \
--run_instance_name your_vis_runVisualization will be saved to:
results/visualization_results/your_vis_run/visualization.png
python main.py --test \
--checkpoint_path results/train_results/your_run_name/madounet_best.pth \
--image_dir /path/to/images \
--mask_dir /path/to/masks \
--run_instance_name test_runResults are saved to:
results/test_results/test_run/test_metrics.csv
python main.py --inference \
--checkpoint_path results/train_results/your_run_name/madounet_best.pth \
--image_dir /path/to/input_images \
--run_instance_name infer_runPredicted masks will be saved under:
results/inference_results/infer_run/
--input_size 256 256— Resize input images.--num_samples 5— For visualization.--threshold 0.5— Binarization threshold for mask predictions.
Feel free to explore or modify the pipeline via main.py for different use-cases.
We gratefully acknowledge the following contributors:
- Kaushal Sambanna — Lead Researcher & Developer
- Sanjana Jhansi Ganji — Researcher, Dataset Preparation, Evaluation Support
- Srikanth Panigrahi — Academic Guidance
- Routhu Srinivasa Rao - Guide
Feel free to open a pull request if you'd like to contribute!
This project is licensed under the MIT License.
You are free to use, modify, and distribute this software with attribution.
For queries, collaborations, or feedback:
- 📧 Email: kaushal.sambanna@gmail.com, sanjanajhansi99@gmail.com
- 🧠 LinkedIn: linkedin.com/in/kaushal-sambanna, linkedin.com/in/sanjana-jhansi-ganji
- 💻 GitHub Issues: Submit an issue
⭐️ If you find this project useful, please consider starring the repository. It helps others discover it!