Path-Seg-Model: Simple Pathology Segmentation

Abstract

A segmentation pipeline for the GlaS MICCAI 2015 Gland Segmentation dataset, featuring a UNet baseline and an extendable interface for transformer-based architectures such as Swin-UNet. The project includes training, evaluation, visualization utilities, and TensorBoard logging.

Hardware Note: Experiments were run on a single NVIDIA H800 80 GB GPU, but the code should run on any CUDA-capable GPU with appropriate batch size adjustments.

🛠 Environment Setup

cd path-seg-model

conda create -n pathseg python=3.10
conda activate pathseg

pip install -r requirements.txt

📁 Project Structure

path-seg-model/
├── main.py                   # Entry point for training / evaluation
├── train.py                  # Training loop logic
├── eval.py                   # Evaluation / inference script
├── requirements.txt          # Python dependencies
├── data/
│   ├── __init__.py
│   └── dataset.py            # Custom Dataset class for GlaS
├── glas/
│   └── Warwick_QU_Dataset/   # Root directory for GlaS data
├── models/
│   ├── __init__.py
│   ├── unet.py               # UNet implementation
│   └── swinunet.py           # Swin-UNet implementation
├── checkpoints/
│   ├── unet_checkpoint.pth
│   └── swinunet_checkpoint.pth
├── utils/
│   ├── __init__.py
│   ├── metrics.py            # Dice, IoU, etc.
│   └── viz.py                # Visualization utilities
└── logs/                     # TensorBoard logs (event files)

📦 Dataset

GlaS MICCAI 2015 Dataset:
https://www.kaggle.com/datasets/sani84/glasmiccai2015-gland-segmentation

Place the extracted dataset in:

glas/Warwick_QU_Dataset/

Expected structure:

glas/
└── Warwick_QU_Dataset/
    ├── Train/
    │   ├── images/*.bmp
    │   └── masks/*_anno.bmp
    └── Test/
        ├── images/*.bmp
        └── masks/*_anno.bmp

🚀 How to Run

Training

python main.py

Optional example:

python main.py --model unet --epochs 100 --batch-size 4

Evaluation / Inference

python eval.py --model unet --checkpoint checkpoints/unet_checkpoint.pth

📊 TensorBoard Monitoring

tensorboard --logdir logs

Open your browser at:
http://localhost:6006

🧠 Models Implemented

• UNet --- models/unet.py\
• Swin-UNet --- models/swinunet.py

Architectures can be extended by adding new files under models/ and integrating them in main.py.

📈 Evaluation Metrics

Currently implemented or planned:

• Dice Coefficient
• *Intersection over Union (IoU)

See utils/metrics.py for implementation details.

📬 Contact

If you encounter any issues, please open a GitHub issue or email:

📧 sylviachung.22@intl.zju.edu.cn