This repository provides the official implementation of MOSAIC, a multi-scale, rotation-aware transformer-based framework designed for robust nuclei segmentation and centroid/heatmap prediction in histopathology images (ER-IHC, MoNuSAC, CoNSeP).

The model integrates hierarchical feature extraction, rotation-aware multi-scale fusion, and transformer-based long-range reasoning, enabling accurate segmentation under scale variation, heterogeneous morphology, and rotational distortions.

• Multi-Scale Patch Extraction

  • Large: 512×512
  • Medium: 256→512
  • Small: 128→512
  • All scales form parallel training streams.

• Rotation-Aware Feature Fusion

  • Designed to maintain robustness across nuclei orientations.

• Transformer Context Aggregation

  • Enhances long-range spatial reasoning.

• Dual Task Prediction

  • Semantic segmentation (multi-class or binary)
  • Centroid/heatmap estimation for instance separation

• Dataset Ready

  • ER-IHC
  • MoNuSAC
  • CoNSeP

├── mosaic.ipynb        # Main notebook: preprocessing, model, training, and evaluation
├── /Dataset/           # Example dataset folders (user should add)
└── README.md

• Python 3.8+
• TensorFlow 2.x / Keras
• NumPy, OpenCV, Matplotlib
• scikit-image, scikit-learn
• tqdm, seaborn

Install using:

pip install tensorflow numpy opencv-python matplotlib scikit-image scikit-learn seaborn tqdm

On Kaggle, all dependencies are already available.

1. Upload or open mosaic.ipynb in a Kaggle Notebook.

2. Add your dataset(s) as input sources.

3. Update dataset paths inside the notebook:

   data_path = "/kaggle/input/your-dataset/"

4. Enable GPU or TPU.

5. Run all cells sequentially:

   • Multi-scale patch generation
   • Label and mask preprocessing
   • Model definition (CNN + Transformer fusion)
   • Training
   • Validation & visualization

The best model is automatically saved to:

/kaggle/working/MOSAIC_savedmodel/

1. Clone the repository:

   git clone https://github.com/<your-repo>/MOSAIC.git
   cd MOSAIC

2. Open Jupyter Notebook:

   jupyter notebook mosaic.ipynb

3. Update dataset paths inside the notebook.

4. Run cells in order.

GPU is strongly recommended for training.

Running MOSAIC will produce:

• Predicted nuclei segmentation masks

• Predicted centroid heatmaps

• Multi-scale visualization plots

• Quantitative metrics:

  • Dice
  • AJI
  • PQ
  • F1 score

• Trained model folder (MOSAIC_savedmodel)

Optional post-processing:

• Watershed-based instance segmentation
• Centroid-guided instance splitting

If you use this framework, preprocessing pipeline, or model structure, please cite the MOSAIC publication (update with your final BibTeX once accepted/published):

@article{yourcitation2025mosaic,
  title={MOSAIC: Multi-Scale Orientation-Aware Segmentation and Instance Classification Network},
  author={Sufyan, Arbab and ...},
  journal={Expert Systems with Applications},
  year={2025},
  publisher={Elsevier}
}