Occult Lymph Node Metastasis Prediction

This repository contains the whole process of our experiments.

1. Environment installation

pip install -r requirements.txt

2. Preprocessing

Put the database in the ./datasets/image_mask_old directory, while place the data information in the ./datasets/training_validation.json, ./datasets/test_1.json and ./datasets/test_2.json files.

Run the following command:

python preprocessing.py

We will obtain the preprocessed dataset in the ./datasets/image_mask_new directory.

The overall structure of the dataset is as follows:

• datasets/
  • image_mask_old/
    • id_1/
      • image.nii.gz
      • mask.nii.gz
    • id_2/
      • image.nii.gz
      • mask.nii.gz
    • id_3/
      • image.nii.gz
      • mask.nii.gz
    • id_4/
      • image.nii.gz
      • mask.nii.gz
    • ...
  • image_mask_new/
    • id_1/
      • image.nii.gz
      • mask.nii.gz
    • id_2/
      • image.nii.gz
      • mask.nii.gz
    • id_3/
      • image.nii.gz
      • mask.nii.gz
    • id_4/
      • image.nii.gz
      • mask.nii.gz
    • ...
  • training_validation.json

    {
        "train": [
            {
                "image": "id_1/image.nii.gz",
                "mask": "id_1/mask.nii.gz",
                "label": label_1,
                "radiomics": [
                    feature_1_1,
                    feature_1_2,
                    feature_1_3,
                    feature_1_4,
                    feature_1_5,
                    feature_1_6,
                    feature_1_7,
                    feature_1_8,
                    feature_1_9,
                    feature_1_10
                ]
            },
            ...
        ],
        "validation": [
            {
                "image": "id_2/image.nii.gz",
                "mask": "id_2/mask.nii.gz",
                "label": label_2,
                "radiomics": [
                    feature_2_1,
                    feature_2_2,
                    feature_2_3,
                    feature_2_4,
                    feature_2_5,
                    feature_2_6,
                    feature_2_7,
                    feature_2_8,
                    feature_2_9,
                    feature_2_10
                ]
            },
            ...
        ]
    }
    

  • test_1.json

    {
        "test": [
            {
                "image": "id_3/image.nii.gz",
                "mask": "id_3/mask.nii.gz",
                "label": label_3,
                "radiomics": [
                    feature_3_1,
                    feature_3_2,
                    feature_3_3,
                    feature_3_4,
                    feature_3_5,
                    feature_3_6,
                    feature_3_7,
                    feature_3_8,
                    feature_3_9,
                    feature_3_10
                ]
            },
            ...
        ]
    }
    

  • test_2.json

    {
        "test": [
            {
                "image": "id_4/image.nii.gz",
                "mask": "id_4/mask.nii.gz",
                "label": label_4,
                "radiomics": [
                    feature_4_1,
                    feature_4_2,
                    feature_4_3,
                    feature_4_4,
                    feature_4_5,
                    feature_4_6,
                    feature_4_7,
                    feature_4_8,
                    feature_4_9,
                    feature_4_10
                ]
            },
            ...
        ]
    }
    

3. Pre-trained Models

We download the pre-trained models of ResNet-18, ResNet-34, ResNet-50 , and Swin Transformer on large-scale medical datasets, and place resnet_18_23dataset.pth, resnet_34_23dataset.pth, resnet_50_23dataset.pth, and model_swinvit.pt files in the ./models directory.

4. Train, Validation and Test

The model is built using PyTorch. All details have been assembled in the ./train and ./test directories. Please replace the "data_dir" parameter in the following .sh files with the absolute path, and then execute the following commands to train, validate, and test the models:

sh train_test_resnet18_radiomics_cat_only.sh

sh train_test_resnet18_pretrain.sh

sh train_test_resnet18_pretrain_radiomics_cat_only.sh

sh train_test_resnet18_pretrain_radiomics_cat.sh

sh train_test_resnet18_pretrain_radiomics_add.sh

sh train_test_resnet18_pretrain_radiomics_lr_add.sh

sh train_test_resnet34_radiomics_cat_only.sh

sh train_test_resnet34_pretrain_radiomics_cat_only.sh

sh train_test_resnet50_radiomics_cat_only.sh

sh train_test_resnet50_pretrain_radiomics_cat_only.sh

sh train_test_densenet_radiomics_cat_only.sh

sh train_test_swintr_radiomics_cat_only.sh

sh train_test_swintr_pretrain_radiomics_cat_only.sh

The results will be saved in the ./output directory.

5. Citation

If you find our work to be useful for your research, please consider citing.

@article{tian2024predicting,
  title={Predicting occult lymph node metastasis in solid-predominantly invasive lung adenocarcinoma across multiple centers using radiomics-deep learning fusion model},
  author={Tian, Weiwei and Yan, Qinqin and Huang, Xinyu and Feng, Rui and Shan, Fei and Geng, Daoying and Zhang, Zhiyong},
  journal={Cancer Imaging},
  volume={24},
  number={1},
  pages={8},
  year={2024},
  publisher={Springer}
}