User login
Conventional chest x-rays combined with artificial intelligence (AI) can identify lung damage from COVID-19 and differentiate coronavirus patients from other patients, improving triage efforts, new research suggests.
The AI tool – developed by Jason Fleischer, PhD, and graduate student Mohammad Tariqul Islam, both from Princeton (N.J.) University – can distinguish COVID-19 patients from those with pneumonia or normal lung tissue with an accuracy of more than 95%.
“We were able to separate the COVID-19 patients with very high fidelity,” Dr. Fleischer said in an interview. “If you give me an x-ray now, I can say with very high confidence whether a patient has COVID-19.”
The diagnostic tool pinpoints patterns on x-ray images that are too subtle for even trained experts to notice. The precision of CT scanning is similar to that of the AI tool, but CT costs much more and has other disadvantages, said Dr. Fleischer, who presented his findings at the virtual European Respiratory Society International Congress 2020.
“CT is more expensive and uses higher doses of radiation,” he said. “Another big thing is that not everyone has tomography facilities – including a lot of rural places and developing countries – so you need something that’s on the spot.”
With machine learning, Dr. Fleischer analyzed 2,300 x-ray images: 1,018 “normal” images from patients who had neither pneumonia nor COVID-19, 1,011 from patients with pneumonia, and 271 from patients with COVID-19.
The AI tool uses a neural network to refine the number and type of lung features being tracked. A UMAP (Uniform Manifold Approximation and Projection) clustering algorithm then looks for similarities and differences in those images, he explained.
“We, as users, knew which type each x-ray was – normal, pneumonia positive, or COVID-19 positive – but the network did not,” he added.
Clinicians have observed two basic types of lung problems in COVID-19 patients: pneumonia that fills lung air sacs with fluid and dangerously low blood-oxygen levels despite nearly normal breathing patterns. Because treatment can vary according to type, it would be beneficial to quickly distinguish between them, Dr. Fleischer said.
The AI tool showed that there is a distinct difference in chest x-rays from pneumonia-positive patients and healthy people, he said. It also demonstrated two distinct clusters of COVID-19–positive chest x-rays: those that looked like pneumonia and those with a more normal presentation.
The fact that “the AI system recognizes something unique in chest x-rays from COVID-19–positive patients” indicates that the computer is able to identify visual markers for coronavirus, he explained. “We currently do not know what these markers are.”
Dr. Fleischer said his goal is not to replace physician decision-making, but to supplement it.
“I’m uncomfortable with having computers make the final decision,” he said. “They often have a narrow focus, whereas doctors have the big picture in mind.”
This AI tool is “very interesting,” especially in the context of expanding AI applications in various specialties, said Thierry Fumeaux, MD, from Nyon (Switzerland) Hospital. Some physicians currently disagree on whether a chest x-ray or CT scan is the better tool to help diagnose COVID-19.
“It seems better than the human eye and brain” to pinpoint COVID-19 lung damage, “so it’s very attractive as a technology,” Dr. Fumeaux said in an interview.
And AI can be used to supplement the efforts of busy and fatigued clinicians who might be stretched thin by large caseloads. “I cannot read 200 chest x-rays in a day, but a computer can do that in 2 minutes,” he said.
But Dr. Fumeaux offered a caveat: “Pattern recognition is promising, but at the moment I’m not aware of papers showing that, by using AI, you’re changing anything in the outcome of a patient.”
Ideally, Dr. Fleischer said he hopes that AI will soon be able to accurately indicate which treatments are most effective for individual COVID-19 patients. And the technology might eventually be used to help with treatment decisions for patients with asthma or chronic obstructive pulmonary disease, he noted.
But he needs more data before results indicate whether a COVID-19 patient would benefit from ventilator support, for example, and the tool can be used more widely. To contribute data or collaborate with Dr. Fleischer’s efforts, contact him.
“Machine learning is all about data, so you can find these correlations,” he said. “It would be nice to be able to use it to reassure a worried patient that their prognosis is good; to say that most of the people with symptoms like yours will be just fine.”
Dr. Fleischer and Dr. Fumeaux have declared no relevant financial relationships.
A version of this article originally appeared on Medscape.com.
Conventional chest x-rays combined with artificial intelligence (AI) can identify lung damage from COVID-19 and differentiate coronavirus patients from other patients, improving triage efforts, new research suggests.
The AI tool – developed by Jason Fleischer, PhD, and graduate student Mohammad Tariqul Islam, both from Princeton (N.J.) University – can distinguish COVID-19 patients from those with pneumonia or normal lung tissue with an accuracy of more than 95%.
“We were able to separate the COVID-19 patients with very high fidelity,” Dr. Fleischer said in an interview. “If you give me an x-ray now, I can say with very high confidence whether a patient has COVID-19.”
The diagnostic tool pinpoints patterns on x-ray images that are too subtle for even trained experts to notice. The precision of CT scanning is similar to that of the AI tool, but CT costs much more and has other disadvantages, said Dr. Fleischer, who presented his findings at the virtual European Respiratory Society International Congress 2020.
“CT is more expensive and uses higher doses of radiation,” he said. “Another big thing is that not everyone has tomography facilities – including a lot of rural places and developing countries – so you need something that’s on the spot.”
With machine learning, Dr. Fleischer analyzed 2,300 x-ray images: 1,018 “normal” images from patients who had neither pneumonia nor COVID-19, 1,011 from patients with pneumonia, and 271 from patients with COVID-19.
The AI tool uses a neural network to refine the number and type of lung features being tracked. A UMAP (Uniform Manifold Approximation and Projection) clustering algorithm then looks for similarities and differences in those images, he explained.
“We, as users, knew which type each x-ray was – normal, pneumonia positive, or COVID-19 positive – but the network did not,” he added.
Clinicians have observed two basic types of lung problems in COVID-19 patients: pneumonia that fills lung air sacs with fluid and dangerously low blood-oxygen levels despite nearly normal breathing patterns. Because treatment can vary according to type, it would be beneficial to quickly distinguish between them, Dr. Fleischer said.
The AI tool showed that there is a distinct difference in chest x-rays from pneumonia-positive patients and healthy people, he said. It also demonstrated two distinct clusters of COVID-19–positive chest x-rays: those that looked like pneumonia and those with a more normal presentation.
The fact that “the AI system recognizes something unique in chest x-rays from COVID-19–positive patients” indicates that the computer is able to identify visual markers for coronavirus, he explained. “We currently do not know what these markers are.”
Dr. Fleischer said his goal is not to replace physician decision-making, but to supplement it.
“I’m uncomfortable with having computers make the final decision,” he said. “They often have a narrow focus, whereas doctors have the big picture in mind.”
This AI tool is “very interesting,” especially in the context of expanding AI applications in various specialties, said Thierry Fumeaux, MD, from Nyon (Switzerland) Hospital. Some physicians currently disagree on whether a chest x-ray or CT scan is the better tool to help diagnose COVID-19.
“It seems better than the human eye and brain” to pinpoint COVID-19 lung damage, “so it’s very attractive as a technology,” Dr. Fumeaux said in an interview.
And AI can be used to supplement the efforts of busy and fatigued clinicians who might be stretched thin by large caseloads. “I cannot read 200 chest x-rays in a day, but a computer can do that in 2 minutes,” he said.
But Dr. Fumeaux offered a caveat: “Pattern recognition is promising, but at the moment I’m not aware of papers showing that, by using AI, you’re changing anything in the outcome of a patient.”
Ideally, Dr. Fleischer said he hopes that AI will soon be able to accurately indicate which treatments are most effective for individual COVID-19 patients. And the technology might eventually be used to help with treatment decisions for patients with asthma or chronic obstructive pulmonary disease, he noted.
But he needs more data before results indicate whether a COVID-19 patient would benefit from ventilator support, for example, and the tool can be used more widely. To contribute data or collaborate with Dr. Fleischer’s efforts, contact him.
“Machine learning is all about data, so you can find these correlations,” he said. “It would be nice to be able to use it to reassure a worried patient that their prognosis is good; to say that most of the people with symptoms like yours will be just fine.”
Dr. Fleischer and Dr. Fumeaux have declared no relevant financial relationships.
A version of this article originally appeared on Medscape.com.
Conventional chest x-rays combined with artificial intelligence (AI) can identify lung damage from COVID-19 and differentiate coronavirus patients from other patients, improving triage efforts, new research suggests.
The AI tool – developed by Jason Fleischer, PhD, and graduate student Mohammad Tariqul Islam, both from Princeton (N.J.) University – can distinguish COVID-19 patients from those with pneumonia or normal lung tissue with an accuracy of more than 95%.
“We were able to separate the COVID-19 patients with very high fidelity,” Dr. Fleischer said in an interview. “If you give me an x-ray now, I can say with very high confidence whether a patient has COVID-19.”
The diagnostic tool pinpoints patterns on x-ray images that are too subtle for even trained experts to notice. The precision of CT scanning is similar to that of the AI tool, but CT costs much more and has other disadvantages, said Dr. Fleischer, who presented his findings at the virtual European Respiratory Society International Congress 2020.
“CT is more expensive and uses higher doses of radiation,” he said. “Another big thing is that not everyone has tomography facilities – including a lot of rural places and developing countries – so you need something that’s on the spot.”
With machine learning, Dr. Fleischer analyzed 2,300 x-ray images: 1,018 “normal” images from patients who had neither pneumonia nor COVID-19, 1,011 from patients with pneumonia, and 271 from patients with COVID-19.
The AI tool uses a neural network to refine the number and type of lung features being tracked. A UMAP (Uniform Manifold Approximation and Projection) clustering algorithm then looks for similarities and differences in those images, he explained.
“We, as users, knew which type each x-ray was – normal, pneumonia positive, or COVID-19 positive – but the network did not,” he added.
Clinicians have observed two basic types of lung problems in COVID-19 patients: pneumonia that fills lung air sacs with fluid and dangerously low blood-oxygen levels despite nearly normal breathing patterns. Because treatment can vary according to type, it would be beneficial to quickly distinguish between them, Dr. Fleischer said.
The AI tool showed that there is a distinct difference in chest x-rays from pneumonia-positive patients and healthy people, he said. It also demonstrated two distinct clusters of COVID-19–positive chest x-rays: those that looked like pneumonia and those with a more normal presentation.
The fact that “the AI system recognizes something unique in chest x-rays from COVID-19–positive patients” indicates that the computer is able to identify visual markers for coronavirus, he explained. “We currently do not know what these markers are.”
Dr. Fleischer said his goal is not to replace physician decision-making, but to supplement it.
“I’m uncomfortable with having computers make the final decision,” he said. “They often have a narrow focus, whereas doctors have the big picture in mind.”
This AI tool is “very interesting,” especially in the context of expanding AI applications in various specialties, said Thierry Fumeaux, MD, from Nyon (Switzerland) Hospital. Some physicians currently disagree on whether a chest x-ray or CT scan is the better tool to help diagnose COVID-19.
“It seems better than the human eye and brain” to pinpoint COVID-19 lung damage, “so it’s very attractive as a technology,” Dr. Fumeaux said in an interview.
And AI can be used to supplement the efforts of busy and fatigued clinicians who might be stretched thin by large caseloads. “I cannot read 200 chest x-rays in a day, but a computer can do that in 2 minutes,” he said.
But Dr. Fumeaux offered a caveat: “Pattern recognition is promising, but at the moment I’m not aware of papers showing that, by using AI, you’re changing anything in the outcome of a patient.”
Ideally, Dr. Fleischer said he hopes that AI will soon be able to accurately indicate which treatments are most effective for individual COVID-19 patients. And the technology might eventually be used to help with treatment decisions for patients with asthma or chronic obstructive pulmonary disease, he noted.
But he needs more data before results indicate whether a COVID-19 patient would benefit from ventilator support, for example, and the tool can be used more widely. To contribute data or collaborate with Dr. Fleischer’s efforts, contact him.
“Machine learning is all about data, so you can find these correlations,” he said. “It would be nice to be able to use it to reassure a worried patient that their prognosis is good; to say that most of the people with symptoms like yours will be just fine.”
Dr. Fleischer and Dr. Fumeaux have declared no relevant financial relationships.
A version of this article originally appeared on Medscape.com.