Many of us think of artificial intelligence (AI) as computers mimicking humans in thinking and behavior. Thankfully, the reality is still very different and AI is more of an evolving tool that can tackle narrow tasks that humans are otherwise not good enough at. One of those tasks is looking at complex images of tissue biopsy samples to find tell-tale signs of the existence of cancers. Pathologists take years to receive enough training to differentiate and identify histologic tissue types, but even those with decades of experience are often stumped when figuring out specific cases.
Now a collaboration of researchers, headed by a team from NYU School of Medicine, have developed an artificial intelligence program that can look at tissue slices of lung tumors and identify their types, including whether a given tumor is expressing specific genes that lead to irregular cellular growth. The researchers were able to fine tune their system so well that it is able to accurately tell adenocarcinoma from squamous cell carcinoma 97% of the time, something that trained human pathologists are still struggling with.
The research team’s platform was Google’s Inception v3, a deep convolutional neural network that was trained to analyze images from the Cancer Genome Atlas, a repository of images of individual cases that have received confirmed diagnoses. The new AI program was able to check its accuracy using known correct answers from the Atlas, allowing it to really optimize its diagnostic prowess.
“In our study, we were excited to improve on pathologist-level accuracies, and to show that AI can discover previously unknown patterns in the visible features of cancer cells and the tissues around them,” said Narges Razavian, PhD, one of the authors of the paper appearing in Nature Medicine. “The synergy between data and computational power is creating unprecedented opportunities to improve both the practice and the science of medicine.”
While the system is able to differentiate the two types of lung cancers, its ability to detect and identify specific gene mutations just from the microscope images of tissue samples of lung cancers is still limited. The researchers are now working on improving the system and training it so it can have an accuracy of at least 90%.
Image: AI tool analyzes a slice of cancerous tissue to create a map that tells apart two lung cancer types, with squamous cell carcinoma in red, lung squamous cell carcinoma in blue, and normal lung tissue in gray.
Study in Nature Medicine: Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning…
Via: NYU…