计算机也能当医生
Computers can analyze chest X-rays almost as well as humans and the two can correct each other's mistakes, a new study shows.
"We basically found that the computer can analyze an image almost as well as a human being," says Dr. James A. Scott, associate professor, Harvard Medical School.
"I'm not saying I would trust a computer to do it in isolation, but with a reasonably well-defined small image like we have in nuclear medicine, a computer was able to interpret it reasonably accurately," Scott says.
, 百拇医药
He was lead author on the study with colleagues, Dr. Edwin L. Palmer and Dr. Alan J. Fischman, at the Massachusetts General Hospital's department of radiology, division of nuclear medicine, and Harvard Medical School.
The three doctors, each with more than 15 years experience in nuclear medicine, examined chest X-rays from 100 patients. The doctors had never seen the images before and were given no details about the patients.
, http://www.100md.com They compared their results to those from computers running neural network software that analyzed the same X-rays.
In most cases, the computers matched the humans in their assessments of the chest scans, particularly in cases with large pulmonary clots. Scott says he suspects the computers would outperform a resident or intern.
In viewing the chest scans, the humans and computers made different types of errors, but a combination of both computer and human diagnosis allowed one to correct the mistakes of the other.
, 百拇医药
The study was published in the August issue of the American Journal of Roentgenology.
Scott stresses the finding doesn't mean computers should be used on their own. But he says that he and his colleagues think computers could be a good training tool or used as a second opinion if a colleague isn't available.
"It doesn't so much improve the diagnostic ability of the human, as it limits the mistakes people make. Because people are going to make mistakes. There's just no way to stop it. What we're trying to do is limit that," Scott says. "If the computer says something 180 degrees to what you thought, it's probably a good idea to revise your opinion and at least say, 'I'm not so sure.' "
, 百拇医药
The computer can't be perfect because humans can give it only so much information, he notes.
"The algorithms we use are just not sophisticated enough to pick up all possible ways of looking at the image. So the computer was kind of limited in how it looked at it," Scott says.
And while humans have the advantage of more thorough knowledge, experience and intuition, their decisions can be influenced by what they read in textbooks, journals and other forms of medical literature.
, 百拇医药
"The computer is not influenced by the literature. In other words, it's not saying, 'I read this from Dr. so-and-so,' or something like that. It's simply looking at the image. And sometimes we'll make mistakes because we just read an article in the literature that said something and maybe the article wasn't right. (The computer's) really looking at it coming from a whole different place than the physician is," Scott says.
Neural network software, which uses pattern recognition to reach its conclusions, is used by a number of people in the financial world to analyze market trends, says Sridhar Narayan, assistant professor of computer science, University of North Carolina.
, 百拇医药
He says neural networks are useful in medicine, as long as they're used carefully.
"I think they will gain acceptance, but they will always be second- or third-fiddle in terms of working in the background and always being doublechecked by some human expert. Mainly because the stakes are so much higher when it comes to using neural networks to do some kind of medical screening," he says.
"We basically found that the computer can analyze an image almost as well as a human being," says Dr. James A. Scott, associate professor, Harvard Medical School.
"I'm not saying I would trust a computer to do it in isolation, but with a reasonably well-defined small image like we have in nuclear medicine, a computer was able to interpret it reasonably accurately," Scott says.
, 百拇医药
He was lead author on the study with colleagues, Dr. Edwin L. Palmer and Dr. Alan J. Fischman, at the Massachusetts General Hospital's department of radiology, division of nuclear medicine, and Harvard Medical School.
The three doctors, each with more than 15 years experience in nuclear medicine, examined chest X-rays from 100 patients. The doctors had never seen the images before and were given no details about the patients.
, http://www.100md.com They compared their results to those from computers running neural network software that analyzed the same X-rays.
In most cases, the computers matched the humans in their assessments of the chest scans, particularly in cases with large pulmonary clots. Scott says he suspects the computers would outperform a resident or intern.
In viewing the chest scans, the humans and computers made different types of errors, but a combination of both computer and human diagnosis allowed one to correct the mistakes of the other.
, 百拇医药
The study was published in the August issue of the American Journal of Roentgenology.
Scott stresses the finding doesn't mean computers should be used on their own. But he says that he and his colleagues think computers could be a good training tool or used as a second opinion if a colleague isn't available.
"It doesn't so much improve the diagnostic ability of the human, as it limits the mistakes people make. Because people are going to make mistakes. There's just no way to stop it. What we're trying to do is limit that," Scott says. "If the computer says something 180 degrees to what you thought, it's probably a good idea to revise your opinion and at least say, 'I'm not so sure.' "
, 百拇医药
The computer can't be perfect because humans can give it only so much information, he notes.
"The algorithms we use are just not sophisticated enough to pick up all possible ways of looking at the image. So the computer was kind of limited in how it looked at it," Scott says.
And while humans have the advantage of more thorough knowledge, experience and intuition, their decisions can be influenced by what they read in textbooks, journals and other forms of medical literature.
, 百拇医药
"The computer is not influenced by the literature. In other words, it's not saying, 'I read this from Dr. so-and-so,' or something like that. It's simply looking at the image. And sometimes we'll make mistakes because we just read an article in the literature that said something and maybe the article wasn't right. (The computer's) really looking at it coming from a whole different place than the physician is," Scott says.
Neural network software, which uses pattern recognition to reach its conclusions, is used by a number of people in the financial world to analyze market trends, says Sridhar Narayan, assistant professor of computer science, University of North Carolina.
, 百拇医药
He says neural networks are useful in medicine, as long as they're used carefully.
"I think they will gain acceptance, but they will always be second- or third-fiddle in terms of working in the background and always being doublechecked by some human expert. Mainly because the stakes are so much higher when it comes to using neural networks to do some kind of medical screening," he says.
健康123网站独家提供，如欲转载，请标明出处
, 百拇医药(Robert Preidt)参见：首页 > 保健版 > 家庭健康 > 电脑保健