Too often, searching for the right cancer treatment is a guess-and-check process, as doctors move through a list of costly therapies that may or may not work until they find one that sticks. But a newly discovered biomarker that can be calculated by an artificial intelligence algorithm from a routine CT scan could one day make that process much more efficient.
The biomarker was identified by researchers representing Emory University, Cleveland Clinic, NYU Langone Health, Weill Cornell Medicine and more. They trained an algorithm to comb through scans of non-small cell lung cancer (NSCLC) tumors, looking for specific indicators of how well an individual might respond to immunotherapy treatments, as described in a retrospective study recently published in Science Advances.
Immunotherapy is often the first treatment option for NSCLC—the most common form of lung cancer—even though it works in less than half of patients and can set them back hundreds of thousands of dollars.
“Immunotherapy only tends to benefit approximately 30% of patients. With the high expense of treatments and a 70% failure rate, we have to find better ways to predict and monitor responses to therapy,” said Anant Madabhushi, Ph.D., an author of the study and a professor in Emory and Georgia Tech’s shared biomedical engineering department.
The new biomarker has been dubbed quantitative vessel tortuosity, or QVT, as it assesses the arrangement of the blood vessels surrounding a tumor. Typically, tumor-associated vasculature is more twisted and chaotically arranged than normal blood vessels, allowing the cancer cells to grow faster and spread more easily throughout the body.
Using AI tools to examine these vessels in more than 500 NSCLC patients both before and after they’d been treated with immune checkpoint inhibitor therapies, the study’s authors determined that tumor-associated vasculature is even more twisted in patients who don’t respond well to immunotherapy—because, they’ve surmised, the more twisted the vessels, the harder it is for anti-tumor cells to actually reach the tumor.
With those findings, the researchers were able to train a machine learning algorithm to analyze the blood vessels around a tumor using only a CT scan and calculate the likelihood that a patient will respond to immunotherapy. The AI was ultimately able to predict treatment response with an average area under the curve of about 0.65 across three test sets, representing an accuracy level slightly above that of random chance, denoted by an AUC of 0.5.
The QVT biomarker and its AI calculator still need to be further developed and validated in clinical trials, but the proof-of-concept results already suggest the tech could be a “game-changer,” according to Mohammadhadi Khorrami, Ph.D., first author on the study and postdoctoral fellow in the joint biomedical engineering department.
“With the staggering costs of immunotherapy—around $200,000 a year per patient—the need to noninvasively determine this response before initiating therapy becomes crucial,” Khorrami said, noting that the biomarker could help doctors offer potentially more effective cancer treatments from the get-go—and keep patients from racking up unnecessarily high hospital bills in the process.
Plus, as Madabhushi added, the QVT approach can assess treatment-induced tumor changes much earlier than other methods, “even before more obvious changes like tumor size become apparent.”