MTMR-Net

Code for Multi-Task Deep Model with Margin Ranking Loss for Lung Nodule Analysis on IEEE Transactions on Medical Imaging (TMI).

Introduction

This repository provides the PyTorch implementation for our TMI paper "Multi-Task Deep Model with Margin Ranking Loss for Lung Nodule Analysis". Our model can output a more robust benign-malignant classification result with persuasive semantic feature scores compared to other CAD techniques which can only output classification results.

Requirement

Python == 2.7.13
PyTorch == 0.3.0
tensorboardX == 0.9
numpy == 1.14.3

Installation

Download and unzip this project:

git clone https://github.com/lihaoliu-cambridge/mtmr-net.git
cd mtmr-net

Download resnet50.pth into ./logs/middle_result_logs/imagenet/ folder.

Dataset

Download the original LIDC-IDRI dataset into ./data/ folder

The preprocessing methods can be found in below two links:
https://github.com/zhwhong/lidc_nodule_detection
https://github.com/jcausey-astate/NoduleX_code

Running

• Modify the args.yaml, add the parameters of your deep learning model under the "running_params" item. Details are shown in deep-learning-model-saving-helper project.

• Pass the running_params (a python dict which contains the running parameters) to you own model.

• The first parameter "is_training" is True for training mode, "is_training" is False for test mode.

• Finish you mode(training or test), and run it.

  cd mtmr-net
  python main.py

Citation

If you use our code for your research, please cite our paper:

@article{liu2019multi,
  title={Multi-task deep model with margin ranking loss for lung nodule analysis},
  author={Liu, Lihao and Dou, Qi and Chen, Hao and Qin, Jing and Heng, Pheng-Ann},
  journal={IEEE transactions on medical imaging},
  volume={39},
  number={3},
  pages={718--728},
  year={2019},
  publisher={IEEE}
}

Question

Please open an issue or email [email protected] for any questions.

Acknowledgement

😙Thanks my dearest brother Yong for this beautiful figure.