DualU-Net is a multi-task deep learning model for nuclei segmentation and classification. It employs a dual-decoder design that predicts both pixel-wise segmentation maps and Gaussian-based centroid density maps, enabling fast and accurate instance segmentation in histopathological images.

Read the MIDL25 paper (Oral Presentation)

Note: This repository is under active development. Please anticipate frequent changes.

DualU-Net has been evaluated on several public datasets:

• PanNuke: A dataset for tissue and nuclei segmentation in histopathology. Link
• CoNSeP: A colorectal cancer nuclei segmentation dataset. Link

The provided PanNuke checkpoints are trained on one fold of the dataset, validated on another, and tested on the third. The naming convention of the checkpoints follows this pattern:

pannuke-combined-{encoder}-{train_fold}{val_fold}.pth

• {encoder}: Specifies the encoder used in the model (e.g., convnext).
• {train_fold}: Indicates the fold used for training.
• {val_fold}: Indicates the fold used for validation.

For example, the checkpoint:

pannuke-combined-convnext-23.pth

• Trained on fold 2
• Validated on fold 3
• Tested on fold 1
• Uses ConvNeXt as the encoder

For CoNSeP, the checkpoint follows a simpler naming convention:

consep-combined-{encoder}.pth

• {encoder}: Specifies the encoder used in the model (e.g., convnext).

For example:

consep-combined.pth

• Trained on the Train CoNSeP dataset
• Uses a combined training approach

1. Clone this repository:

   git clone https://github.com/davidanglada/DualU-Net.git

2. Create a virtual environment:

   python -m venv dualunet-env
   source dualunet-env/bin/activate 

3. Install required dependencies:

   pip install -r requirements.txt

DualU-Net/
├── dual_unet/                         # Main module containing core functionalities
│   ├── datasets/                      # Dataset building from COCO, transforms and augmentation
│   ├── eval/                          # Detection and segmentation evaluation functions.
│   ├── models/                        # Contains model architectures and related utilities
│   ├── utils/                         # Contains utility functions
│   ├── __init__.py                    # Initialization file for the dual_unet module
│   └── engine.py                      # Train and evaluation functions
├── configs/                           # Configuration files
│   ├── train_config.yaml              # Configuration file for training
│   ├── eval_config.yaml               # Configuration file for evaluation
├── eval.py                            # Script for evaluating the model
├── train.py                           # Script for training the model
├── requirements.txt                   # List of required Python packages
└── README.md                          # Project documentation

Below are the basic commands to get you started with DualU-Net.

python train.py --config configs/config_train.yaml

python eval.py --config configs/config_eval.yaml

python inference.py --config configs/config_inference.yaml

If you find this work helpful in your research, please consider citing us:

   @inproceedings{
      anglada-rotger2025two,
      title={Two Heads Are Enough: DualU-Net, a Fast and Efficient Architecture for Nuclei Instance Segmentation},
      author={David Anglada-Rotger and Berta Jansat and Ferran Marques and Montse Pard{\`a}s},
      booktitle={Medical Imaging with Deep Learning},
      year={2025},
      url={https://openreview.net/forum?id=lK0CklgxQd}
      }

This project (all code and non-code assets) is released under the
Creative Commons Attribution – NonCommercial 4.0 International License.