Platform enables AI to sift pathology images to ID differences in lung cancer cells
Researchers at two companies say they have developed an artificial intelligence platform intended to aid pathologists who use images to classify subtypes of lung cancer.
NantHealth and NantOmics say they have developed an artificial intelligence-based machine learning digital pathology software that can be used by pathologists to identify tumor-infiltrating killer cells, called lymphocytes, from whole slide images.
The companies say findings of initial tests demonstrate novel AI methods of identifying tumor-infiltrating lymphocytes (TILs) in lung cancers.
Researchers shared an initial report on the technology at the recent international joint conference of the American Association for Cancer Research and the International Association for the Study of Lung Cancer. The study concluded that despite lower overall tumor mutation burden and lymphocyte levels, there exists a subset of lung cancers with very high infiltrating lymphocyte counts.
Findings are important because it demonstrates the way technology can be used to personalize treatments for patients, researchers say. The inclusion of AI can speed the process and improve patient results, they add.
“Accurately identifying and quantifying tumor-infiltrating white cells is extremely important for prognosis and treatment decisions in this era of personalized medicine, yet it currently requires manual review of whole slide images by medically trained pathologists, and incurs significant delays and cost,” says Patrick Soon-Shiong, MD, chairman and CEO of NantHealth. “Our goal was to develop a scalable remote cloud-based diagnostic imaging system. To accomplish this, machine vision of digitally transmitted images of tumor tissue would facilitate a scalable cloud-based infrastructure, with an image patch-based, automated system to classify cancers by their immune status.”
Use of the technology can enable highly accurate tumor-region and lymphocyte detection, perhaps enabling oncologists to use the information to better treat patients with the most appropriate treatment for their version of cancer cells.
The system was trained and tested on 876 subtyped NSCLC gigapixel-resolution diagnostic whole slide images from 805 patients obtained from The Cancer Genome Atlas sources. Samples were randomly split into sets for training (711 whole slide images from 664 patients) and testing (165 whole slide images from 141 patients).
Findings show that NantOmics and NantHealth’s fully automated histopathology subtyping AI method outperforms other algorithms reported in literature for diagnostic whole slide images. The system also generated maps of (tumor) regions-of-interest within whole slide images, providing novel spatial information on tumor organization.