Researchers create an AI model that reliably forecasts the course of cancer

The study demonstrates AI's capacity to understand the spatial relationships between cells within tissues, extracting nuanced information that was previously incomprehensible to humans.

An important step forward in the use of AI for personalized treatment plans and the likely course of cancer patients' disease has been made by researchers who have created an artificial intelligence model that reliably predicts patient outcomes from tissue samples.

According to the researchers, cell spatial organization is similar to an intricate jigsaw puzzle in which each cell is a distinct piece that must fit together precisely to form a coherent tissue or organ structure. (Rutgers Cancer on Twitter)

The journal Nature Communications describes an innovative approach that examines the spatial arrangement of cells in tissue samples.

According to the researchers, cell spatial organization is similar to an intricate jigsaw puzzle in which each cell is a distinct piece that must fit together precisely to form a coherent tissue or organ structure.

Guanghua Xiao, a professor at the University of Texas Southwestern Medical Center in the US, led the study. "The study showcases the remarkable ability of AI to grasp these intricate spatial relationships among cells within tissues, extracting subtle information previously beyond human comprehension while predicting patient outcomes," Xiao said.

Patients' tissue samples are routinely taken, and pathologists examine and interpret them on slides in order to diagnose patients.

However, the researchers noted that interpretations can differ among pathologists and that this process is time-consuming.

Furthermore, they noted, the human brain is capable of overlooking minute details found in pathology images that could be crucial indicators of a patient's state.

A pathologist's duties can be partially completed by a variety of AI models developed in recent years, such as recognizing different cell types or utilizing cell proximity as a stand-in for interactions between cells.

Nevertheless, more intricate aspects of pathologists' interpretation of tissue images—like identifying patterns in the spatial organization of cells and eliminating "noise" in images that could confuse interpretations—are not well captured by these models.

The newly developed artificial intelligence model, Ceograph, begins by identifying cells in pictures and their locations, just like pathologists read tissue slides.

From there, it distinguishes between different cell types and maps their morphology and spatial distribution to produce a map that allows for the analysis of cell interactions, arrangement, and distribution.

Using pathology slides, the researchers were able to successfully apply this tool to three clinical scenarios.

In one, they used Ceograph to differentiate between squamous cell carcinoma and adenocarcinoma, two subtypes of lung cancer.

In a second, they forecasted the chance that oral abnormalities, which are precancerous lesions of the mouth, could develop into cancer.

The third involved determining which patients with lung cancer had the best chance of responding to a particular class of drugs known as epidermal growth factor receptor inhibitors.

The Ceograph model significantly outperformed conventional techniques in each scenario when it came to predicting patient outcomes.

Significantly, according to Xiao, the cell spatial organization features found by Ceograph are interpretable and provide biological insights into how changes in a single cell-cell's spatial interaction might have a variety of functional effects.

He continued, "These results point to an expanding role for AI in medical care, providing a means of enhancing the effectiveness and precision of pathology analyses."

According to Xiao, "This approach has the potential to optimize treatment selection for individual patients and streamline targeted preventive measures for high-risk populations."