RCC-T3a_Invasion-radiomics

License

This code is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Introduction

This is the official repository for our recent work: [CT-based radiomic model for identifying pathological T3a upstaging in renal cell carcinoma: model development and multi-source validation]. This repository contains ready-to-use codes for inference of the proposed models.

Key Features

• Features 1: The training dataset contains sufficient patients with renal cell carcinoma (n = 999) and covers a wide spectrum of disease phenotype （common and rare RCC subtypes, all ISUP/fuhrman grades, all pathological T stages）.
• Features 2: The predictive performance of the models were externally evaluated in data from two medical centers and four TCIA datasets.
• Features 3: The morphology model improved the performance of junior radiologists.

Get Started

Main Requirements

numpy
pandas
pyradiomics
SimpleITK
scikit-learn
tqdm

Data preparation

NIfTI-formatted image and mask files were used for feature extraction and model prediction. Please organize the image and mask files in the following structure.

test_data/
├── Dataset001
    │── patient_1
    │    ├── image.nii.gz
    │    ├── tumor_mask.nii.gz
    │    └── peritumor_mask.nii.gz
    │── patient_2
    │    ├── image.nii.gz
    │    ├── tumor_mask.nii.gz
    │    └── peritumor_mask.nii.gz
    └── patient_n
         ├── image.nii.gz
         ├── tumor_mask.nii.gz
         └── peritumor_mask.nii.gz

Config preparation

During inference, the configurations are loaded from a .json file. The structure of the file is as follows:

{
    "model":"your\\path\\to\\model.pkl", 
    "extraction parameter":"your\\path\\to\\extraction.yaml", 
    "dataset":"your\\path\\to\\dataset", 
    "image":"image.nii.gz", 
    "masks": {
        "tumor":"tumor_mask.nii.gz", 
        "peritumor":"peritumor_mask.nii.gz"
    }, 
    "cutoff":0.3582773836548992, 
    "output":"your\\path\\to\\results\\"
}

Each of the parameters is explained as follows:

• model: The path to the pickle file of trained model.
• extraction parameter: The path to the yaml file of radiomics feature extraction configurations.
• dataset: The path to the dataset folder.
• image: The name of the image file. Both .nii and .nii.gz are supported.
• masks: The names of the mask files. The key is the prefix added to the feature name during extraction, and the value is the mask file name.
• cutoff: The cutoff value that determines whether a case is classified as T3a invasion positive or negative.
• output: The output folder to hold the output .csv file.

Inference

from inference import ModelInference

if __name__ == '__main__':
    config_dir = '.\\inference_config.json'
    model = ModelInference(config_dir)
    model.predict()

1. import ModelInference class from inference module.
2. initialize an instance of the ModelInference class, passing the configuration file path (inference_config.json) as an argument.
3. call the predict method of the ModelInference class.

Model Training and validation

1. use SignatureExtractor class in inference module and .yaml configuration file to extract the radiomic features.
2. merge the tabular radiomics data with label column and save it as .csv file.
3. run train in train_validation module to train the model and save it as .pkl file.
4. run internal validaion in train_validation module to internal validate the model using nested-cross validation. The outer loop is 200 repeats 5 fold cross validation, and the inner loop is 20 repeats 5 fold cross validation. For a detailed explanation of nested-cross validation, please refer to A Guide to Cross-Validation for Artificial Intelligence in Medical Imaging.
5. run external_validation in train_validation module to external validate the model in holdout datasets.
6. during internal validation and external validation, the raw predictions were saved as .csv files.