Synthetic data could be key to AI applications in medicine

Applications of AI

Advances in artificial intelligence (AI) have revolutionized many scientific fields, including medicine. However, this technique poses data acquisition problems. AI needs annotated data to learn and ultimately perform with sufficient accuracy, but high-quality data is often not easily accessible, such as in complex and novel surgical scenarios. .

A researcher from the Laboratory for Computational Sensing and Robotics has published a paper. nature machine intelligence We described a new model that can create realistic synthetic X-ray image data for training AI algorithms.

Computer-assisted surgery is a broad class of techniques that help surgeons plan various surgical interventions. The researchers in this paper were particularly interested in improving surgical navigation techniques that would allow surgeons to track the position of surgical instruments on patient images taken before surgery.

Cong Gao, the lead author of the paper and a 2022 graduate of the University’s Computer Science PhD program, explained the research in the following interview: Newsletter.

“If you look at a self-driving car, navigation is wanting to know where the car is on the map,” he said. “In a surgical scenario, the map is your body. Navigation here mainly refers to how to provide the surgeon with enough information so that they can place the surgical tools in their planned positions. “

AI algorithms can be used to determine where surgical instruments are relative to different landmarks on 2D X-ray images. We use data from both surgical instruments and the patient’s x-rays to train our AI model, but it’s difficult to get enough of these very specific images. Instead, researchers tried to create synthetic data in simulations.

A major obstacle in creating synthetic data is that its characteristics may differ from those of the actual patient data. AI trained on synthetic data may only work on other synthetic images, which creates problems when predictions need to be made on real data. This problem is called the domain gap.

However, researchers were able to get around this by taking synthetic data and making major changes to the images using a concept called domain randomization. Adding these modified images to the training set can force the network to recognize new patterns and improve performance. With this concept they developed his SyntheX. It is a system for creating AI X-ray imaging algorithms trained only on synthetic data that perform as well as or better than AI trained on real data.

The researchers compared the synthetic data model to the real data model in three example tasks: hip joint imaging, surgical instrument detection, and a COVID-19 chest x-ray dataset. Controlling all other variables, they took x-ray images of him in real life and computed tomography (CT) scans of 366 corpses and recreated composite x-ray images in simulations replicating the real data. bottom.

AI models trained on synthetic data were able to detect specific structures and landmarks in patient X-rays with similar error rates, and in some cases outperformed real data models. The potential for synthetic data from AI to be used in surgery creates opportunities for surgeons and researchers.

Gao explained how one application is creating AI systems for new surgical instruments that have not yet been tested on patients.

“You can also use this tool to generate scenarios that you haven’t run yet. Let’s say you have a new surgical tool that you want to test, but you don’t have a real dataset yet,” he said. “Plugging a patient’s CT scan of her and a model of surgical instruments into this algorithm produces a highly realistic X-ray image of him that includes both objects.”

Another important application is SyntheX’s ability to create large datasets for AI training. AI tends to be more accurate the larger the dataset it trains on.

In the hip imaging task, SyntheX was able to create 10,000 composite images from just 20 hip CT scans. This allows researchers to avoid the major bottleneck of obtaining large datasets from real patient populations.

Additionally, the creation of synthetic data eliminates the ethical and privacy concerns that arise when obtaining patient data for AI. This allows researchers to reduce development time for many innovative surgical applications.

Gao is researching new uses for AI with soft tissue applications, and plans to work on more advanced systems of human modeling in the future. He hopes the technology will help surgeons and improve patient outcomes.

“If we can envision an entire surgical procedure in a purely simulated environment, we can anticipate many obstacles before entering the operating room,” he said.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *