Curtis Langlotz, MD, PhD, Professor of Radiology and Biomedical Informatics Research, Dr. Akshay Chaudhari, Assistant Professor of Radiology and Biomedical Data Science, has developed Roentgen, an open AI model that generates realistic synthetic X-rays from medical X-rays.
Transcript:
Langlotz:
There is a big gap in medical machine learning right now. This is a rare condition, especially when there are rare diseases. There is not enough data to train these models. Composite data becomes a big part of the puzzle.
Sergios Gatidis:
The first distal phalanx-based dysplastic appearance, the fourth central eye head, and the fourth and fifth metacalpal metaphysics.
Langlotz:
When studying AI in the 1980s, it could take four years and a PhD, and we would develop a system that could work with three or four patients. Today, you can build a system in a week with the most accurate accuracy of the time, using the right training data. The fact that radiology has now been a digital specialist for nearly 20 years is one of the reasons why I think radiology is leading the way in AI.
Chaudhari:
One of the biggest challenges facing radiological AI research is that there is usually no large datasets to train AI models. To be able to train AI models, you must always curate large, high-quality datasets. High quality can be simple, but for the most part it can be a little difficult. Maybe I need to aggregate around 1 million chest X-rays, but what if I only have 100,000 images? How can you build a high quality AI model if you have a small dataset?
Langlotz:
Some students saw what was happening outside of medicine using a model that allows them to create images based on text prompts. Show off your kangaroo and Ray-Ban sunglasses in Rembrandt style. That way you'll draw a picture. I wonder what happens if I ask any of these models. Their return was really like a chest x-ray cartoon. It didn't look like a chest x-ray, and these students readjusted the model.
Chaudhari:
Immediately after seeing that tool, I was blown away. Can you convert this toy into a useful resource for researchers and convert it into a toy instead of training an AI model with actual data? Can you create some of these composite images that are similar to what the actual image looks like? And that's why the Roentgen model was born. Let's see how this model actually works. Let's start with this image of a dog. Next, add a little noise to this. This noise follows a specific pattern we pre-calculated. What we do is train a simple model to remove to return to the original image. You can then take this noisy image and then add a little more noise. I'm actually using the same removal model to remove this very loud image and try to return to the original clean image. It continues to add more and more noise. Reaching images with only noise. You can start with a random noise distribution and take a step back and run the removal model, but instead of using a dog, you are actually using chest x-rays. We also use radiation reporting information to actually guide you how this removal process works.
Langlotz:
The purpose of Roentgen is to create additional data used to train AI software tools to provide additional accuracy to radiologists working in the clinic, identify diseases faster, identify diseases that may be missed, and extend them to other areas such as CTS and MRI and ultrasound.
Chaudhari:
Roentgen can be used to create a composite image of what a patient will look like if he or she contracts pneumonia. Another research group may be interested in trying to identify cardiac tumors, which are cardiac enlargements. We just want to be the tide where we can lift all the boats. We want to be able to create these high quality datasets for all downstream tasks.
Langlotz:
RoentGen can be used to reduce bias and implements the algorithm by creating synthetic data for several subgroups that do not have sufficient training data. That means improving patient privacy. This means a more accurate AI model and a more responsible implementation of AI algorithms.
Stefania Moroianu:
Hey everyone. One thing I wanted to share today about Roentgen V2 is this interactive reasoning demo. Here you can get a closer look at how the model responds to various prompts. One cool update for Roentgen V2 and V1 is the introduction of this demographic information at a prompt that can condition the patient's age, race and gender. For example, we show a normal chest x-ray for a man. When you change to a woman, you will notice visual changes and chest x-rays that correspond to women.
Chaudhari:
Stanford is trying to build sophisticated AI algorithms to handle healthcare data. So I'm really excited to be able to do this research and understand what your capabilities are. I've started to realize that all different applications can have. How can you analyze the contents of these images? Can you predict future illnesses that a patient may have? Can a radiologist help write a radiologist's report? At least in large-scale imaging models, that seems like an imminent future. The possibilities of these models are combined with doing radiology to make them advantageous for both parties.
Langlotz:
While the open source software movement is extremely important and likely to be the future of AI, there is a greater chance that it will actually increase the pace of progress, build these tools faster and innovate faster.
Chaudhari:
And at the end of the day, it's not just about using the model at Stanford. We hope that the model will support patients around the world. So, if we can open up some of the tools we are building, we hope it will help us translate better healthcare solutions.
Details: Synthetic data
