Do you have models trained and/or evaluated on Chest ImaGenome?

[PabloMessina]

Hi, just a very quick question. Chest ImaGenome (https://physionet.org/content/chest-imagenome/1.0.0/) provides very fine-grained labels and bounding boxes for most images in MIMIC-CXR. Do you guys have models trained and/or evaluated using this dataset?

Best regards,
Pablo

[ieee8023]

Currently no. But it looks like a very awesome dataset that we should have. Here is a notebook showing the current datasets with mask information: https://github.com/mlmed/torchxrayvision/blob/master/scripts/xray_masks.ipynb

[PabloMessina]

That's awesome. You already support several datasets with mask and bounding box annotations. The cool thing about Chest ImaGenome is that it comes with bounding boxes for 36 different anatomical locations + very fine-grained scene graphs describing frontal chest X-ray images, for 240K+ images of MIMIC-CXR, so it's a very large scale dataset. You can read more about how they developed the dataset here: https://arxiv.org/pdf/2108.00316.pdf

Here are some papers that have already used Chest ImaGenome, to give you an idea of the things that can be done with it:

• AnaXNet: Anatomy Aware Multi-label Finding Classification in Chest X-ray: https://arxiv.org/pdf/2105.09937.pdf
• Anatomy Aware Model for Longitudinal Relationships Change in CXRs: https://arxiv.org/pdf/2208.03873.pdf
• Anatomy-Guided Weakly-Supervised Abnormality Localization in Chest X-rays: https://arxiv.org/pdf/2206.12704.pdf
• Few-shot Structured Radiology Report Generation Using Natural Language Prompts: https://arxiv.org/pdf/2203.15723.pdf

[ieee8023]

Thanks for that info! Do you know how to work with the data already? Can you help me get started with a dataset that loads the masks similar to the existing datasets to prepare a PR?

[PabloMessina]

I've been playing around with the dataset for a while, but I have my own ad-hoc customized ways to load and post-process the data. I can point you to specific sections of my code if that helps though. For example:

• Scripts to read and compute post-processed versions of the annotations: https://github.com/PabloMessina/MedVQA/tree/master/medvqa/scripts/chest_imagenome
• A bunch of utility functions to work with Chest ImaGenome: https://github.com/PabloMessina/MedVQA/tree/master/medvqa/datasets/chest_imagenome
• Here I combine data from MIMIC-CXR with data from Chest ImaGenome to create a dataset and a dataloader (plus a bunch of other things). Not the prettiest code: https://github.com/PabloMessina/MedVQA/blob/master/medvqa/datasets/mimiccxr/mimiccxr_vision_dataset_management.py

[PabloMessina]

This jupyter notebook might be helpful as well: https://github.com/PabloMessina/MedVQA/blob/master/medvqa/datasets/notebooks/Exploring%20Chest%20ImaGenome.ipynb (Note: it's a work in progress)

[PabloMessina]

Hey @ieee8023, just so you know, there is another paper recently published in CVPR 2023 using the Chest ImaGenome dataset: Interactive and Explainable Region-guided Radiology Report Generation:

• paper: https://arxiv.org/pdf/2304.08295.pdf
• code: https://github.com/ttanida/rgrg