> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# MIT's FleXray outlines 60 anatomical structures on standard X-rays
- URL: https://globalfeed.ai/en/mits-flexray-outlines-60-anatomical-structures-on-standard-x-rays/
- Published: 2026-09-25T21:10:16.000Z
- Updated: 2026-09-25T21:10:16.000Z
- Description: Researchers from MIT, Massachusetts General Hospital and Harvard Medical School have released FleXray, an open model that automatically outlines organs and bones on X-rays. It was trained on synthetic X-rays made from CT scans and can be tried in the browser without the image leaving your computer.
- Author: GlobalFeed Editor
- Tags: MIT, medical imaging, X-ray, open models, x-MIT_CSAIL, x-ion_barrel, dil-en, elle

Researchers from **MIT** CSAIL, the Martinos Center at Massachusetts General Hospital and Harvard Medical School have released **FleXray**, a model that automatically outlines anatomy on X-ray images. It can mark 60 anatomical structures on X-rays from different body regions, angles and machines. The code, model files and training data are all public.

## Why this is hard

According to the project page, more than 2.3 billion X-rays are taken every year, yet very little data exists that marks organ boundaries pixel by pixel. An X-ray squeezes the 3D body into one 2D image, so organs overlap, boundaries disappear and the same region looks very different from different angles. That is why earlier models were usually trained for one region or one view.

## How it was trained

The researchers worked around the data gap with CT scans. From 3D CT volumes where organs are already labelled, they rendered synthetic X-rays from a different angle at every training step, turning one patient's scan into a stream of distinct X-ray and label pairs. Because these renders look too clean, the team used Black Forest Labs' FLUX.2 image editing model to add film grain, exposure differences and burned-in text like real X-rays, without changing the anatomy. Real X-rays were also part of training.

## Results

According to the paper, FleXray was tested on eight real X-ray datasets it never saw during training, covering hand and wrist, elbow, leg, hip, spine, ribs and adult and paediatric chest. It beat the generalist models it was compared with and came close to models trained specifically on each dataset. The researchers also tried it on three further tasks:

- **Scoliosis measurement:** It calculated the Cobb angle of spinal curvature without task-specific training. According to the paper, its error was lower than the typical difference between two doctors measuring the same X-ray.
- **Surgical image alignment:** It sharply cut the failure rate when aligning a pre-operative CT with live X-ray during a procedure.
- **New tasks with little data:** It adapted to new targets such as hip dysplasia in children and bone tumours with only a few labelled examples.

## How to try it

On FleXray's project page you can upload an X-ray and run the model directly in your browser; according to the page, the image never leaves your computer. It can also be installed as a Python package, with code on GitHub and model files and training data on Hugging Face. The team also built an MCP server to connect the model to AI assistants.

The work comes from Victor Ion Butoi, Vivek Gopalakrishnan, John Guttag, Adrian Dalca and Neel Dey at MIT. The researchers stress that the demo is a research tool, not a medical device, and must not be used for diagnosis. The code is MIT-licensed, while the model files are released under CC BY-NC 4.0, which allows non-commercial use only. The results were shared in a pre-print; to see how the model performs in clinical settings, it is worth following evaluations from independent researchers.

Sources: [FleXray project page](https://flexray.csail.mit.edu/?ref=globalfeed.ai), [arXiv: FleXray paper](https://arxiv.org/abs/2609.26756?ref=globalfeed.ai), [GitHub: FleXray](https://github.com/VictorButoi/FleXray?ref=globalfeed.ai), [Hugging Face: model files](https://huggingface.co/VictorButoi/flexray?ref=globalfeed.ai).