2️⃣ Our prior study linking muscle quality & quantity to hundreds of diseases:
doi.org/10.1016/j.eb...
3️⃣ Our prior work predicting future heart disease using body composition (and more):
doi.org/10.1038/s415...
4️⃣ Our open-source body composition segmentation library:
github.com/stanfordmimi...

See more details (including why I started lifting at the gym in part because of this research) here: tinyurl.com/aictsarcopenia

🔗 and some other links to dig deeper:
1️⃣ Our new study on detecting low muscle mass with AI from CT imaging:
pubs.rsna.org/doi/10.1148/...

aand of course I lost track of the n/n count. We really need better ways to share longer ideas.

📄 Learn more:
🔗 LLaVA-Rad MIMIC-CXR Annotations: physionet.org/content/llav...
🔗 LLaVA-Rad preprint: arxiv.org/abs/2403.08002
🔗 CheXprompt: github.com/microsoft/ch... (our validated, radiologist-aligned metric for assessing report quality)
Peer-reviewed manuscript, code/model weights: coming soon ;)

We call this resource the LLaVA-Rad MIMIC-CXR Annotations. In the spirit of reproducibility and collaboration, we are releasing the full training and evaluation dataset to support further innovation.

1️⃣ No references to prior imaging: We removed mentions of previous X-rays to reduce the risk of models generating incorrect descriptions tied to unavailable prior exams.

2️⃣ Cleaner, section-specific extractions: We refined the extraction of report sections and descriptions.

3/n

[...] describing chest X-rays in the manner of a radiologist. One major challenge we encountered was the lack of publicly available, high-quality, and standardized data.

To address this, we built on the MIMIC-CXR dataset and leveraged GPT-4 to create annotations that improve in two key ways:

2/n