A.I. Harnessed to Detect Cancer Cells

(NIH.gov)

10 Dec. 2019. A software package using an image-aided form of machine learning can quickly identify and classify tumor from other cells in lung cancer images. Researchers from Southwestern Medical Center in Dallas, part of the University of Texas system, report their findings in the 22 November issue of the journal EBioMedicine.

A team led by bioinformatics professor Guanghua Xiao is seeking faster methods for detecting signs of cancer in digital pathology images, which are now examined by human experts. As a result, say the authors from Southwestern’s Quantitative Biomedical Research Center, pathologists examine details from representative sections, not the entire the image, but still a time-consuming and sometimes error-prone process.

“As there are usually millions of cells in a tissue sample, a pathologist can only analyze so many slides in a day,” says Xiao in a Southwestern statement. “To make a diagnosis, pathologists usually only examine several ‘representative’ regions in detail, rather than the whole slide. However, some important details could be missed by this approach.”

For example, says Xiao, even expert eyes and minds can inadvertently miss subtle changes in patterns, making it difficult to spot changes in tumors and their support system, called the tumor microenvironment. To meet these needs, the researchers developed software that examines digital pathology slides, automatically classifies tumor and other cells, and quantifies their distribution on the slide.

The heart of the software, called ConvPath, is a convolutional neural network, a form of artificial intelligence. A convolutional neural network combines image analysis and machine learning to dissect an image by layers for understanding features in the image. Different aspects of each layer discovered and analyzed by the algorithm are translated into data that the algorithm then uses to train its understanding of the problem being solved, with that understanding enhanced and refined as more images and data are encountered.

The ConvPath algorithms are trained with digital pathology slides and related clinical data from 523 lung cancer patients in an image database maintained by The Cancer Genome Atlas, a project of National Cancer Institute and National Human Genome Research Institute, divisions of NIH. In practice, pathologists first identify regions of interest in slides, with identifiable signs of cancer, then the software identifies the nucleus in the cells and places the nuclei in sample patches as small as 80 by 80 pixels from the digital images.

From there, ConvPath analyzes each slide image layer by layer, identifying cells as tumors, stromal or connective tissue, or lymphocytes, a form of white blood cells. The software then classifies and maps the interactions of the cells, and quantifies their distribution throughout the image space.

The researchers tested and validated ConvPath with separate collections of patient data from The Cancer Genome Atlas and a similar patient database in China. The results show classification accuracy of 90 to 93 percent during the training and testing phases of the project. The team then used ConvPath data to develop a prognosis for lung cancer patients, calculating risks among patients for demographic factors, such as age and sex, as well as smoking status and progression of their lung cancer that generally correlated to the patients’ clinical outcomes.

The authors make the software and test data available to the public from the Southwestern web site.

More from Science & Enterprise:

• App Detects Early Eye Disease Symptoms from Photos
• A.I. Finding Cancer Treatments in Generic Drugs
• Chip Device Performs Detailed Cancer Cell Analysis
• Blood Test as Good as Biopsy to Detect Cancer Targets
• Organoids Designed for Faster Cancer Diagnostics

*     *     *