Clinical AI developed by Harvard University works on par with human radiologists

artificial intelligence concept

The new tool uses natural language descriptions from accompanying clinical reports to identify diseases in chest X-ray images.

New tool overcomes major hurdle in clinical AI design.

Scientists from Harvard Medical School And the Stanford University They have created an AI diagnostic tool that can detect diseases with chest X-rays based on natural language descriptions provided in accompanying clinical reports.

Since most current AI models need tedious human explanation of huge amounts of data before giving out the data labeled in the model to train, the step is a major advance in clinical AI design.

The model, called CheXzero, has been made on a par with human radiologists in its ability to identify diseases in chest X-rays, according to a paper describing their work which is published in The nature of biomedical engineering. The group also made models cipher Openly available to other researchers.

To correctly detect diseases during ‘training’, the majority of AI algorithms need labeled data sets. Because this annotation-intensive, often costly, and time-consuming procedure is required by human clinicians, it is particularly challenging for tasks that involve interpretation of medical images.

For example, to label a chest x-ray data set, radiologists would have to look at hundreds of thousands of x-ray images one by one and comment on each one clearly with the conditions detected. While newer AI models have attempted to address this labeling bottleneck by learning from unlabeled data in the “pre-training” stage, they ultimately require fine-tuning of labeled data to achieve high performance.

By contrast, the new paradigm is self-supervised, meaning that it learns more independently, without the need for manually labeled data before or after training. The form is based solely on chest x-ray images and English notes found in accompanying x-ray reports.

“We are living in the early days of next-generation medical AI models capable of performing flexible tasks through direct learning from text,” said study lead author Pranav Rajpurkar, associate professor of biomedical informatics at the Blavatnik Institute at HMS. “Until now, most AI models have relied on manual annotations for huge amounts of data – up to 100,000 images – to achieve high performance. Our method does not need such disease-specific annotations.

“With CheXzero, one can simply feed the model a chest X-ray and report the corresponding radiograph, and they will learn that the image and text in the report should be considered the same—in other words, they learn to match the chest X-rays with their accompanying report,” Rajpurkar added. How the concepts in the unstructured text correspond to the patterns visible in the image.”

The model was ‘trained’ on a publicly available data set containing more than 377,000 chest radiographs and more than 227,000 corresponding clinical observations. Its performance was then tested on two separate data sets of chest X-ray images and corresponding observations collected from two different institutions, one of which was in a different country. The goal of this diversity of data sets was to ensure that the model performed equally well when exposed to clinical observations that might use different terms to describe the same outcome.

Upon testing, CheXzero successfully identified diseases that were not clearly explained by human physicians. It has outperformed other self-supervised AI tools and performed with[{” attribute=””>accuracy similar to that of human radiologists.

The approach, the researchers said, could eventually be applied to imaging modalities well beyond X-rays, including CT scans, MRIs, and echocardiograms.

“CheXzero shows that accuracy of complex medical image interpretation no longer needs to remain at the mercy of large labeled datasets,” said study co-first author Ekin Tiu, an undergraduate student at Stanford and a visiting researcher at HMS. “We use chest X-rays as a driving example, but in reality, CheXzero’s capability is generalizable to a vast array of medical settings where unstructured data is the norm, and precisely embodies the promise of bypassing the large-scale labeling bottleneck that has plagued the field of medical machine learning.”

Reference: “Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning” by Ekin Tiu, Ellie Talius, Pujan Patel, Curtis P. Langlotz, Andrew Y. Ng, and Pranav Rajpurkar, 15 September 2022, Nature Biomedical Engineering.
DOI: 10.1038/s41551-022-00936-9

Leave a Comment