Joint learning of chest radiographs and radiology reports in the application of pulmonary edema assessment.
This repository incorporates the algorithms presented in
G. Chauhan*, R. Liao* et al. Joint Modeling of Chest Radiographs and Radiology Reports for Pulmonary Edema Assessment. International Conference on Medical Image Computing and Computer-Assisted Intervention, 2020.
(* indicates equal contributions)
Set up the conda environment using conda_environment.yml. You might want to remove the pip dependencies if that is creating an issue for you. You can manually install the spacy and scispacy dependencies using pip install spacy and pip install scispacy. Read https://allenai.github.io/scispacy/ for more information on scispacy.
Train the model in an unsupervised fashion, i.e., only the first term in Eq (3) is optimized:
python ${repo_path}/scripts/main.py
--img_data_dir ${repo_path}/example_data/images/
--text_data_dir ${repo_path}/example_data/text/
--data_split_path ${repo_path}/example_data/data_split.csv
--use_text_data_dir
--use_data_split_path
--output_dir ${output_path}
--do_train
--training_folds 1 2 3 4 5 6
--training_mode 'semisupervised_phase1'
Train the model in a supervised fashion:
python ${repo_path}/scripts/main.py
--img_data_dir ${repo_path}/example_data/images/
--text_data_dir ${repo_path}/example_data/text/
--data_split_path ${repo_path}/example_data/data_split.csv
--use_text_data_dir
--use_data_split_path
--output_dir ${output_path}
--do_train
--training_folds 1 2 3 4 5
--training_mode 'supervised'
Note that in our data split, fold 6 is the unlabeled image-text pairs.
We have experimented this algorithm on MIMIC-CXR, which is a large publicly available dataset of chest radiographs in DICOM format with free-text radiology reports. The dataset contains 377,110 images corresponding to 227,835 radiographic studies performed at the Beth Israel Deaconess Medical Center in Boston, MA.
We have demonstrated the application of this algorithm in pulmonary edema assessment. We aim to classify a given chest x-ray image into one of the four ordinal levels: no edema (0), vascular congestion (1), interstitial edema (2), and alveolar edema (3).
We extract the Impressions, Findings, Conclusion, and Recommendation sections from the radiology reports. If none of these sections are present in the report, we use the Final Report section. We tokenize the text using ScispaCy before providing it to the BERT model. The data pre-processing code is available here. We're releasing the extracted report sections soon!
We use regex to extract pulmonary edema severity labels from the radiology reports for our model training. A board-certified radiologist and two domain experts reviewed and labeled 485 radiology reports (corrsponsding to 531 chest radiographs). We use the expert labels for our model testing. The regex labeling results and expert labels on MIMIC-CXR are summerized here.
In our MICCAI 2020 work, we split the MIMIC-CXR data into training and test sets. There is no patient overlap between the training set and the test set. Our data split can be found here. The folds 1-6 are the training set and the fold "TEST" is the test set. We also used the training set for cross-validation when tuning our model hyper-parameters.
Geeticka Chauhan: geeticka [at] mit.edu
Ruizhi Liao: ruizhi [at] mit.edu