[ad_1]
OOne of the most lethal attributes of lung cancer is its ability to deceive radiologists. Some nodules seem threatening but turn out to be false positives. Others completely escape notification and then wrap up without symptoms in metastatic disease.
Monday, however, Google has unveiled an artificial intelligence system that, in preliminary tests, demonstrated a remarkable talent for detecting disguises of lung cancer.
A study published in Nature Medicine revealed that the algorithm, trained on 42,000 CT scans of patients taken in a clinical trial from the National Institutes of Health, outperformed six radiologists to determine whether patients had cancer. It has detected 5% more cancers and eliminated false positives – when cancer is suspected while a nodule is harmless – by 11% on reading a single scan. The results were comparable to those of radiologists when previous images of patients were also included in the assessment.
publicity
Results highlight the potential of AI to improve lung cancer screening and help radiologists diagnose malignancies earlier and more accurately – although research has not shown whether the Google system would help patients to live longer. Among cancers, lung cancer is by far the most deadly among Americans, making about 160,000 deaths in 2018.
A large study conducted by the NIH – the same one that provided data for Google's algorithm – established that screening high-risk patients, such as long-term smokers, can reduce the risk of death for many people. 39, about 20%, but he also raised concerns about unnecessary false positives tests that may harm patients. He reported that several deaths were due to false positives that led patients to undergo invasive biopsies and other procedures. Persistent concerns about the accuracy and overall benefits of screening have led to low rates of such tests. At the same time, about 50% of lung cancers are diagnosed after their spread, while the disease is more difficult to treat effectively.
Experts who did not participate in the study said Google's system could make testing much more viable.
"These people have a technology that will dramatically improve the accuracy of screening," said Dr. Otis Brawley, professor of oncology and epidemiology at Johns Hopkins University and former executive vice president of the American Cancer Society. He is generally skeptical about lung cancer screening, but points out that Google's performance in reducing false positives is a significant step forward.
"This will prevent other problems for people who get tested," said Brawley, adding that the system's high performance in this regard does not necessarily mean that it will save more lives from lung cancer.
Google's system will require more rigorous testing – probably a randomized controlled trial – before it can be put into medical practice. Since the study is limited to patients already treated, it is therefore impossible to say whether the system, when used on new patients, will result in more effective care and better results.
Google executives acknowledged this fact in the study and an accompanying blog post, adding that they were working with clinical partners to refine and validate the system. "To fully assess this situation, you need to work with research organizations and conduct large-scale trials to understand how this technology will work on a large scale and across large populations," said Daniel Tse, Google Product Manager. study. He added that the company had pre-submission discussions with the Food and Drug Administration to discuss the criteria for approval.
Google engineers who developed the artificial intelligence system pointed out that it was not designed to replace radiologists, but to improve their ability to detect nodules and determine s & # 39; They were dangerous. Existing computer-assisted systems separate the detection and diagnosis of nodules into different tasks. The Google system performs both of these functions, targeting regions of interest in a scan and providing a risk score indicating whether a patient's nodules are cancerous.
The system uses convolutional neural networks, a type of artificial intelligence architecture, to understand the characteristics of malignancy and indicate problem areas by analyzing three-dimensional CT scans. This task is difficult and time consuming for radiologists because they can not examine three-dimensional scans as a computer. they must examine hundreds of individual slices of analysis to solve problems. But the computer can review all dimensions at the same time.
"We were able to train [AI models] very high resolution, "said Shravya Shetty, Google engineer and technical leader of the study. "Although radiologists can examine sections, the model has clear advantages."
The authors reported that system performance remained consistent when exposed to patients outside the NIH dataset on which it was trained. The system examined the analyzes of 1,700 patients at Northwestern Memorial HealthCare in Chicago and produced similar results by classifying nodules and establishing diagnoses.
According to experts, the Google software could be particularly useful for general radiologists, who often review patients' lung scans in US hospitals in the United States.
"To make screening accessible to everyone, it can only be done by chest radiologists. This must be done by all radiologists, "said Dr. Jorge Gomez, a medical oncologist from Mount Sinai Health System, who is the national spokesperson for the American Lung Association.
At Mount Sinai, he said, chest radiologists regularly attend meetings to discuss patient care and highlight specific aspects of scans that indicate why nodules may be malignant or not.
"It's an incredible resource, and computers could do it," he said. "This is a very important study that should entice a rich person to do a randomized trial."
Tse, Google's product manager, said the company was actively pursuing its work. "We are getting things done both internally at Google and with our partners," he said, adding that he was not yet able to set the schedule for the next one. follow-up study.
[ad_2]
Source link