Advanced imaging technology that uses artificial intelligence (AI) can potentially predict which patients with lung cancer are likely to experience cancer progression after surgery, according to new data.
The technology, known as highly multiplexed imaging mass cytometry (IMC), can provide cellular-level detail of the tumor immune microenvironment, which may allow clinicians to identify patients who need additional treatment, as well as those who don’t.
“It is well known that the frequency of certain cell populations within the tumor microenvironment correlates with clinical outcomes. These observations help us understand the biology underlying cancer progression,” senior author Logan Walsh, PhD, assistant professor of human genetics and the Rosalind Goodman Chair in Lung Cancer Research at McGill University’s Rosalind and Morris Goodman Cancer Institute in Montreal, told Medscape Medical News.
“We wanted to test whether using completely unbiased AI could find and use the spatial topography of the tumor microenvironment from IMC data to predict clinical outcomes,” he said. “It turns out the answer is yes! AI can predict clinical outcomes when combined with IMC with extremely high accuracy from a single 1-mm2 tumor core.”
The study was published on February 1 in Nature.
The Immune Landscape
Lung cancer is the leading cause of cancer-related death in Canada, surpassing breast, colon, and prostate cancer deaths combined, the study authors write.
Lung adenocarcinoma, a non–small cell lung cancer, is the most common subtype and is characterized by distinct cellular and molecular features. The tumor immune microenvironment influences disease progression and therapy response, the authors write. Understanding the spatial landscape of the microenvironment could provide insight into disease progression, therapeutic vulnerabilities, and biomarkers of response to existing treatments.
In a collaborative study, Walsh and colleagues from McGill University and Université Laval profiled the cellular composition and spatial organization of the tumor immune microenvironment in tumors from 416 patients with lung adenocarcinoma across five histologic patterns. They used IMC to assess at samples from the universities’ biobanks that patients had provided for research purposes.
The research team detected more than 1.6 million cells, which allowed spatial analysis of immune lineages and activation states with distinct clinical correlates, including survival. They used a supervised lineage assignment approach to classify 14 distinct immune cell populations, along with tumor cells and endothelial cells.
High-grade solid tumors had the greatest immune infiltrate (44.6%), compared with micropapillary (37%), acinar (39.7%), papillary (32.8%), and lepidic architectures (32.7%). Macrophages were the most frequent cell population in the tumor immune microenvironment, representing 12.3% of total cells and 34.1% of immune cells.
The prevalence of CD163+ macrophages was strongly correlated with FOXP3+ immunoregulatory T cells in the solid pattern. This relationship was less pronounced in low-grade lepidic and papillary architectures. This finding could suggest an interplay between macrophage and T-cell populations in the tumor immune microenvironment across lung adenocarcinoma patterns.
Using a deep neural network model, the researchers also analyzed the relationship between immune populations and clinical or pathologic variables by examining the frequency of individual cell types as a percentage of total cells in each image. Each image was cross-referenced with clinical data from patients, including sex, age, body mass index, smoking status, stage, progression, survival, and histologic subtype.
Overall, the researchers found that various clinical outcomes, including cancer progression, could be predicted with high accuracy using a single 1-mm2 tumor core. For instance, they could predict progression in stage IA and IB resected lung cancer with 95.9% accuracy.
“We were not surprised that AI was able to predict clinical outcomes, but we were surprised that it was able to do so with such high accuracy and precision,” said Walsh. “We were also surprised to learn that our predictions were equally accurate using only six-plex data, compared with 35-plex. This hinted to us that we could potentially scale down the number of markers to a practical number that would be amenable to technologies available in routine pathology labs.”
Walsh and colleagues are now validating the predictive tool using a lower-plex technology. In addition, they are investigating the immune landscapes of primary and metastatic brain tumors.
“This study is important, as it helps us to understand and appreciate the biological and mechanistic factors that may influence treatment outcomes. Our standard clinical predictors for predicting risk of recurrence and probability of response to therapy are not optimal,” Yee Ung, MD, an associate professor of radiation oncology at Sunnybrook Health Sciences Centre in Toronto, told Medscape.
Ung, who wasn’t involved with this study, has researched noninvasive hypoxia imaging and targeting in lung cancer. Ideally, he said, future studies should incorporate the use of noninvasive imaging predictive factors, in addition to the tumor immune microenvironment and clinical factors, to predict outcomes and provide personalized treatment.
“As we begin to investigate and understand more about cancer biology down to the cellular and molecular level, we need to strategically use AI methodologies in the processing and analysis of data,” he said.
The study was supported by the McGill Interdisciplinary Initiative in Infection and Immunity, the Brain Tumour Funders’ Collaborative, the Canadian Institutes of Health Research, and the Canadian Foundation for Innovation. Walsh and Ung have disclosed no relevant financial relationships.
Nature. Published February 1, 2023. Full text
Carolyn Crist is a health and medical journalist who reports on the latest studies for Medscape, MDedge, and WebMD.
For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.