Mini-version of the Synthetic X-Rays dataset (300)
#3
by VirginiaFGCarmena - opened
This dataset is a balanced subset of a larger CXR dataset that was generated as follows:
- An expert in radiology came up with a set of positive/negative prompts for pleural effusion, edema and normal lungs
- These prompts were used as input conditioning to the CXR latent diffusion model available in the MONAI model zoo (https://github.com/Project-MONAI/model-zoo/tree/dev/models/cxr_image_synthesis_latent_diffusion_model)
- The resulting PNG files were converted into Dicom files by creating synthetic metadata
- The dicoms were structured into accession ID folders
- A dataframe with the minimum amount of metadata was created to be used to test FLIP: accession_id, and "Yes", "No" labels for all three categories, as it would come out with a valid cohort query on the OMOP CDM with the medical imaging extension database
- A balanced subset of 300 images was created (300 ensures that we don't get NaN on classification metrics due to lack of label representation on train/val/test sets)