FDG-PET/CT_AI

This repository is created to release the model weights for the following paper: "Development and validation of pan-cancer lesion segmentation AI-model for whole-body 18F-FDG PET/CT in diverse clinical cohorts". Please cite the paper if you are using the model.

We have used 3D full resolution nnUNet framework. Follow the instrustion below to run inference on new dataset using our model.

1. First, create a conda environment. You can name it to your liking; for example, 'petct-env'.
2. Install nnUNet. Installation process can be found in the following link: documentation/installation_instructions.md
3. For model development: AUTOPET open access dataset has been used. The dataset can be downloaded from TCIA website
4. Create a new folder (anyname). Create the following three sub-folder in this directory: 'nnUNet_raw', 'nnUNet_preprocessed','nnUNet_results' inside the folder. The names should be exactly the same.
5. Create another folder "Dataset101_PETCT" inside 'nnUNet_raw', 'nnUNet_preprocessed','nnUNet_results' folders. This is important for nnUNet to identify which dataset to process.
6. nnU-Net expects datasets in a structured format. This format is inspired by the data structure of the Medical Segmentation Decthlon. Please read the following link for dataset conversion: how-to-use-nnUNet
7. Image file should be in nifti format. USe the following package: TCIA_processing and use the following command:

  python3 -W ignore tcia_dicom_to_nifti.py /PATH/TO/DICOM/FDG-PET-CT-Lesions/ /PATH/TO/NIFTI/FDG-PET-CT-Lesions/

9. PET images should be renamed as channel 1 input with '_0000.nii.gz' extension and CT images '_0001.nii.gz'. Example PET image: PETCT_0ea07b421b_0000.nii.gz, CT Image: PETCT_0ea07b421b_0001.nii.gz
10. The PET/CT image files needs to be put inside the '/nnUNet_raw/Dataset101_PETCT/imagesTe' path.
11. "dataset_fingerprint.json", "nnUNetPlans.json","dataset.json" files should place inside "/nnUNet_preprocessed/Dataset101_PETCT" path.
12. Model weights can be obtained by request only
13. Plase the model weights inside the following path: "nnUNet_results\Dataset101_PETCT\nnUNetTrainer__nnUNetPlans__3d_fullres/". Inside 'nnUNetTrainer__nnUNetPlans__3d_fullres' folder, model weights from 5 folds in speparate folder be present.
14. Once everything is set, run the bash file "inference.sh" to run inference using the model weights. Please modify the folder paths 'nnUNet_raw', 'nnUNet_preprocessed','nnUNet_results' according to your set up directories inside the .sh file.

##MONAI UNet Training

We have provided scripts to train and run inference MONAI UNet model.

1. Train_monai_Unet.py is used to train the model. Dataset folder, Output folder path need to be set in the script. The MONAI transforms will be varied based on different varient studied in the paper. Modifications are needed to be made accordingly.
2. create_dataset_monai_unet.py file prepares the dataset in MONAI readable format. MONAI expects train data in a dictionary format, with image and label names. This script creats, stratified train, validation and test splis of the AUTOPET dataset and then create dictionary out of this for model training.
3. Inference_monai_Unet.py is used to run inference on validation and test set to observe the performance. Dataset folder, Output folder path need to be set in the script. This script will save the input image, ground truth mask and predicted masks in the output folders for both validation and test set.