Enhancing NSCLC Histological Subtype Classification: A Federated Learning Approach Using Triplet Loss

Usage

Preprocessing

Preprocess the original images

python preprocess.py -mp /path/to/metadata -dp /path/to/original/data -op /path/to/output/folder -ts target_spacing

/path/to/metadata is the path to a CSV file with at least three columns 'Subject ID', 'Modality', and 'File Location'. 'File Location' should be relative to /path/to/original/data. 'Modality' should be 'CT' for the images and 'RTSTRUCT' for segmentation.

target_spacing should be a tuple of floats, default is (0.977, 0.977).

The preprocessed files will be saved with the name <Subject ID>_<slice_idx>.nrrd to the /path/to/output/folder. Additionally, another file will be written to the same output folder with the name center_info.csv containing the center points of tumor bounding boxes.

Fold preparation

Create folds in a stratified fashion with the augmentations for train, val and test sets.

python prepare_folds.py -dn dataset_name -dp /path/to/preprocessed/data -lp /path/to/label/data -cp /path/to/clinical/data -op /path/to/output/folder -dl desired_len

dataset_name should be 'NSCLC-Radiomics' for NSCLC-Radiomics dataset. It is not necessary for others.

/path/to/preprocessed/data is the path to the preprocessed images.

/path/to/label/data is the path to the CSV file with at least two columns 'X' and 'y' for subject ids and labels.

/path/to/clinical/data is the path to the clinical data file (required for NSCLC-Radiomics).

/path/to/output/folder is the folder path to the output Excel file with several sheets with the following names : Fold<fold_no>_<phase>

desired_len is the desired number of instances for one class after the augmentations.

Running experiments

python main.py -rp /paths/to/preprocesed/data -fp /paths/to/fold/files -resp result_file_name -exp exp_id -mode triplet/softmax -freeze

/paths/to/preprocessed/data should have the paths to the folders of preprocessed folders for all datasets. Paths should be seperated with a whitespace.

/paths/to/fold/files should have the paths to the Excel files with the folds. Paths should be seperated with a whitespace. The file should have several sheets with the following names : Fold<fold_no>_<phase>

The results of the experiment will be saved to ./results/<result_file_name>.csv

mode It should be either 'triplet' or 'softmax'. If it is 'triplet', the model will be trained with triplet loss. If it is 'softmax', the model will be trained with softmax loss.

freeze is an optional argument. If it is used, some layers of the model will be frozen.

Other parameters are set as default to the values used in the paper.