[ad_1]
How an AI tool analyzes a slice of cancer tissue to create a map that distinguishes two types of lung cancer, squamous cell carcinoma in red, squamous cell carcinoma of the lung in blue, and normal lung tissue in gray. Credit: NYU School of Medicine
A new study shows that a new computer program can analyze images of patients' lung tumors, specify types of cancer, and even identify altered genes leading to abnormal cell growth.
Led by researchers from the NYU School of Medicine and published online in Medicine of natureThe study found that a type of artificial intelligence program (AI), or "machine learning", could distinguish with an accuracy of 97% adenocarcinoma and squamous cell carcinoma, two types of lung cancer that experienced pathologists sometimes struggle to test without confirmatory tests.
The AI tool was also able to determine whether abnormal versions of 6 lung cancer genes – including EGFR, KRAS and TP53 – were present in the cells, with accuracy varying between 73 and 86% according to the gene. Such genetic changes or mutations often cause the abnormal growth observed in cancer, but can also alter the shape and interactions of a cell with its environment, providing visual cues for automated analysis.
According to the researchers, it is essential to determine which genes are modified in each tumor, thanks to the increased use of targeted therapies against cancer cells with specific mutations. Thus, about 20% of patients with adenocarcinoma have mutations in the epidermal growth factor receptor or EGFR, which can now be treated with approved drugs.
But genetic tests currently used to confirm the presence of mutations may take weeks to get results, say the study's authors.
"Delaying the start of a cancer treatment is never a good thing," says lead author of the study, Aristotelis Tsirigos, Ph.D., associate professor in the department Pathology of the Perlmutter Cancer Center of the University of New York at Langone. "Our study provides strong evidence that an AI approach will be able to instantly determine the cancer subtype and the mutational profile so that patients begin to follow targeted therapies faster."
Automatic learning
In this study, the research team designed statistical techniques that allowed their program to learn how to perform a task better, but without knowing exactly how. Such programs build rules and mathematical models that allow decisions to be made based on sample data, as the program becomes "smarter" as the amount of training data increases.
New approaches to AI, inspired by networks of nerve cells in the brain, use increasingly complex circuits to process information in layers, each step fueling the information and assigning more or less importance to each information.
The current team has formed a deep convolutional neural network, Google's Inception v3, to analyze slide images obtained from the Atlas of the Cancer Genome, a database of data containing images of previously determined cancer diagnoses. This allowed researchers to measure how well their program could be trained to accurately and automatically classify normal versus diseased tissue.
Interestingly, the study found that about half of the low percentage of poorly classified tumor images by the AI program in the study were also poorly classified by pathologists, highlighting the difficulty of distinguishing between the two types of lung cancer. By the way, the machine learning program attributed to 45 of the 54 images incorrectly classified by at least one pathologist in the study, suggesting that the IA could give a second helpful opinion.
"In our study, we were excited to improve the accuracy of pathologists and to show that AI can discover unknown patterns in the visible characteristics of cancer cells and tissues that surround them, "says Narges Razavian, Ph.D., assistant professor in the Departments of Radiology and Population Health. "The synergy between data and computing power creates unprecedented opportunities to improve both the practice and science of medicine."
To go forward, the team plans to continue training its AI program until it can determine which genes are mutated in a given cancer with more than 90% accuracy. the diagnosis of several types of cancer.
Explore more:
Unveiled mutations predisposing lung cancers to refractory histological transformation
More information:
Nicolas Coudray et al, Classification and prediction of mutation from histopathological images of non-small cell lung cancer by deep learning, Medicine of nature (2018). DOI: 10.1038 / s41591-018-0177-5
Source link
