Researchers from UT Southwestern (Dallas, TX,
USA) have developed a software tool that uses artificial intelligence (AI) to
recognize cancer cells from digital pathology images, allowing clinicians to
predict patient outcomes.
The spatial distribution of different types of cells can
reveal a cancer’s growth pattern, its relationship with the surrounding
microenvironment, and the body’s immune response. However, the process of
manually identifying all the cells in a pathology slide is extremely labor
intensive and error-prone. A major technical challenge in systematically
studying the tumor microenvironment is how to automatically classify different
types of cells and quantify their spatial distributions.
The AI algorithm, called ConvPath, developed by the
researchers overcomes these obstacles by using AI to classify cell types from
lung cancer pathology images. The ConvPath algorithm can “look” at cells and
identify their types based on their appearance in the pathology images using an
AI algorithm that learns from human pathologists. The algorithm effectively
converts a pathology image into a “map” that displays the spatial distributions
and interactions of tumor cells, stromal cells (i.e., the connective tissue
cells), and lymphocytes (i.e., the white blood cells) in tumor tissue. Whether
tumor cells cluster well together or spread into stromal lymph nodes is a
factor revealing the body’s immune response. So knowing that information can
help doctors customize treatment plans and pinpoint the right immunotherapy.
Ultimately, the algorithm helps pathologists obtain the most accurate cancer
cell analysis – in a much faster way.
“As there are usually millions of cells in a tissue sample, a
pathologist can only analyze so many slides in a day. 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,” said Dr. Guanghua “Andy” Xiao, corresponding author of a study
published in EBioMedicine and Professor of Population and Data Sciences at UT
Southwestern. “It is time-consuming and difficult for pathologists to locate
very small tumor regions in tissue images, so this could greatly reduce the
time that pathologists need to spend on each image.”
