This module encompasses the training and inference scripts for the text classifier model. It serves the purpose of extracting abnormality labels from datasets through partial manual annotations. The manual annotations utilized for training and evaluating the model are accessible in CT-RATE.
Please follow the installation of CT-CLIP.
Adjust the datast folder and include train.csv and val.csv for your downloaded dataset. Subsequently, execute the following command:
$ python train.py --dataset 'path_to/dataset_folder' --augment 0 --scheduler 'None'
# dataset: specify dataset_folder
# augment: random suffle augmentation. 0 or 1
# scheduler: add scheduler. 'cawr' for cosine annealing warmup, 'rlop' for ReduceLrOnPlatauAdjust the file paths within the bash command to point to test_all.csv (which contains all accessions and reports) and text_transformer_model.pth for your downloaded dataset and model. Subsequently, execute the following command:
$ python infer.py --checkpoint 'path_to/text_transformer_model.pth' --single_data 'path_to/single_csv_file' --save_path 'path_to_save_dir'or
$ python infer.py --checkpoint 'path_to/text_transformer_model.pth' --dataset 'path_to/dataset_folder' --save_path 'path_to_save_dir'# checkpoint: pretrained model weights
# dataset: specify dataset_folder
# single_data: specify single csv fileTo evaluate the model, please run the inference in the validation data. After doing that, adjust the ground truth and inferred outputs in the eval.py script and run it as:
$ python eval.py