Using artificial intelligence (AI), European researchers have developed an algorithm that they say successfully detects molecular changes in tumor cells and tissues from microscopic slides in many different cancers.
In a study, published in Nature Cancer, the researchers found that the algorithm could distinguish between cancerous and noncancerous tissues, as well as identify specific aberrant genomic patterns in 28 different cancer types, from point mutations in driver genes to copy number variations and whole genome duplications (Nat Cancer 2020; https://doi.org/10.1038/s43018-020-0085-8).
“What is quite remarkable is that our algorithm can automatically link the histological appearance of almost any tumor with a very broad set of molecular characteristics and with patient survival,” said Moritz Gerstung, PhD, group leader at EMBL European Bioformatics Institute.
Institute researchers collaborated on the study with scientists from the Wellcome Sanger Institute and Addenbrooke's Hospital in Cambridge, UK.
The pan-cancer analysis is believed to be the largest to date to train computer vision to “see” and combine digital pathology with the genetic changes that occur in cells as malignancies take hold. Ordinarily, histopathologists examine the appearance of cancer tissue under a microscope first, then geneticists perform molecular sequencing separately to analyze changes in the genetic code.
For this study, the researchers repurposed an algorithm developed by Google in 2016 to identify everyday objects from the Internet. Because of its superior performance, the scientists hypothesized it would work equally well in identifying genomic changes in cancers, according to Yu Fu, PhD, a postdoctoral fellow with the Gerstung group at the informatics institute.
Altogether, they analyzed fresh frozen tissue slides collected from the Cancer Genome Atlas in 10,452 individuals with cancer, while 14 normal tissue samples without genomic alterations served as the study's control. The algorithm accurately identified more than 160 DNA patterns and thousands of RNA changes in tumors, as well as both the number and spatial location of tumor-infiltrating lymphocytes in all 28 cancer types.
The TIL finding, in particular, holds considerable short-term promise for improving personalized cancer treatment tailored to the individual, according to Fu. The additional information provided could have an important impact in both the understanding of the tumor microenvironment and on prognosis, especially in the case of immunotherapy, she wrote, in an email response to questions.
Future of AI
Whether AI might some day replace molecular sequencing, however, seems unlikely, except perhaps in the developing world, where sequencing costs are prohibitive and the technology is often unavailable, Fu and others noted. And far more research is needed before this imaging approach moves into the clinical arena.
If translated into clinical practice, it's absolutely essential that researchers move to paraffin-fixed tissues, as fresh frozen tissue is unavailable in 99.9 percent of cancer patients, said Kenneth Aldape, MD, Chief of the Pathology Laboratory at the National Cancer Institute. But, even extending the algorithm's format in this way would not necessarily yield clinically useful information for all cancers.
“Though impressive,” he said, describing the European work, “it only looks at a small proportion of the diversity of cancers in humans.” So, it would be applicable to only these cancer types and not others.
His colleague, Eytan Ruppin, MD, PhD, Chief of the NCI's Cancer Data Science Laboratory, agreed.
“It's a well done study that lays the groundwork for future advancement,” he said. “It's interesting and significant, but the findings are of moderate predictive accuracy. Much more work is needed.”
And, while AI “can really do wonders” in specific domains through visual learning and the ability to assess large amounts of samples, Ruppin noted, to suggest it might some day replace physicians is “quite far from the foreseeable horizon.”
For now, AI's role in cancer, in general, is still evolving, scientists agree. The main application today involves algorithms that help clinicians detect cancerous lesions from imaging technologies—CT scans, MRIs, and X-rays. “I believe it's achieving clinical utility in these areas,” he said.
Nevertheless, relative to AI's diverse and broad applications elsewhere, its tools are under evaluation in health care on a “very small scale,” according to Constance Lehman, MD, PhD, Director of Breast Imaging at Massachusetts General Hospital in Boston and Professor of Radiology at Harvard Medical School. “It's not seen in clinical care where we'd like it to be,” she said.
At her facility, Lehman is developing AI to aid in interpreting mammograms to predict current and future risks for breast cancer.
“For this paper, as well as my own research, we use computer vision to train a machine how to see images,” she noted, specifically to measure breast density, a known risk factor for the disease. Although she envisions AI's future use as primarily synergistic with other technologies, she's clearly impressed by what massive computer-generated data can do.
“AI can amplify weak signals and find subtle cues that the human brain is not picking up. And yes, it can lead to far more accuracy in diagnosis,” Lehman noted.
Susan Jenks is a contributing writer.