RATCHET works as a baseline to be compared with our method. In this git repo you could find the pretrained model and predict free-text reports from X-ray images.
Once you have the free-text reports, you need the pretrained RadGraph model in RadGraph Benchmark to convert the free-text reports to structured graph json files.
It is not enough to have just the RadGraph-format json files, you need following script to build the final dataset.
OBS_ANAT_list.json lists all selected observations and anatomies. gen_dataset.py generates the dataset used in our model, while the input is the radgraph-format json file. It converts all entities names into non-sensitive numbers to input the model.
To save the loading resources, we pre-transform the images and save them as tensors locally. You could find more details in the script file