MTGT

About

MTGT (Multiscale Text Feature-Guided Transformer) is a deep learning framework for medical image segmentation. It uses multiscale text features to guide a Transformer architecture and improve segmentation performance on medical imaging datasets.

Installation

Software Requirements

• Python 3.x

• PyTorch (GPU version recommended for training)

• Common Python libraries used by the repository (for example NumPy, pandas, OpenCV)

Hardware Requirements

• A CUDA-capable GPU with sufficient memory is recommended

• Default batch size is 4; reduce to 2 in Config.py if out-of-memory errors occur

Obtain MTGT

Clone the repository:

git clone https://github.com/zlxokok/MTGT.git

Enter the project directory:

cd MTGT

Usage

MTGT includes model training (main_MTGT.py) and inference (infer_MTGT.py).

Input Files

• Medical image datasets with paired segmentation masks

• Config.py for dataset paths and hyperparameters

• BUSI split files such as Train_text.xlsx and Val_text.xlsx

Training

python main_MTGT.py

Main outputs:

• Trained model checkpoint

• Training logs and metrics (for example loss and mIoU)

Inference

python infer_MTGT.py

Main outputs:

• Predicted segmentation masks

• Evaluation metrics (for example mDice, mIoU, Recall, Precision, and F1-score)

License

This project is publicly shared without a specified license. Contact the repository author for usage permission details.

Contact

For questions about MTGT:

zhaoxuanlong254@gmail.com

Citation

If you use MTGT in your research, cite:

Zhao L, Wang T, Zhang X, et al. MTGT: Multiscale Text Feature-Guided Transformer in medical image segmentation. Image and Vision Computing, 2026, 165: 105846.