NEW YORK – Following validation of an artificial intelligence-powered image-analysis method that can detect glioblastoma progression following treatment, University of Pennsylvania researchers are in discussions with the US Food and Drug Administration to develop a commercial test based on the technology.
The standard treatment for glioblastoma is adjuvant therapy with the chemotherapy drug temozolomide plus radiation based on results of a 2005 study in which median survival for patients on chemotherapy and radiation was 14.6 months compared with 12.1 months for patients treated with radiation alone.
Since then, there have been no major advances in treatments that have changed the prognosis for glioblastoma patients. However, researchers believe their chances could be improved by distinguishing patients with tumor progression from those with pseudoprogression, a type of lesion that looks like a tumor but is a result of treatment with temozolomide. Pseudoprogression occurs in up to 66 percent of glioblastoma patients receiving standard therapy, and those patients are typically able to continue on temozolomide with frequent follow-up MRI scans, while patients with true tumor progression could benefit from additional surgery or switching to another therapy.
Pseudoprogression can occur up to six months after treatment, and conventional MRI scans don't adequately distinguish it from tumor progression. This represents an area of unmet need that can be filled with an AI-driven multiparametric image-analysis tool, Suyash Mohan, a neuroradiologist at Penn Medicine, argued in a publication in Clinical Cancer Research in July 2023. In addition to giving patients an opportunity to switch to a more effective therapy as early as possible, he said such a tool could also provide a useful surrogate endpoint for interventional clinical trials in glioblastoma, leading to an earlier readout than survival-based endpoints.
"When somebody is diagnosed with glioblastoma, it is like a death sentence," Mohan said. "Once you treat somebody with standard-of-care [therapy], recurrence is inevitable, and when this recurrence happens, it is too late already for anybody to do anything."
Mohan and his colleagues developed a tool that combines two MRI techniques — diffusion tensor imaging and dynamic susceptibility contrast-perfusion weighted imaging — to create an AI model for differentiating pseudoprogression from tumor progression in glioblastoma patients who had received surgery followed by chemotherapy and radiation. Mohan said the model essentially assists the radiologist by picking up finer, more granular details from the images that the human eye can't detect.
Mohan views the method as a combination of machine and human intelligence rather than simply an AI tool, since it still depends on human input to interpret the changes on the MRI scans. "This is still AI, but it's augmented intelligence, which means that we are not just using machines, we are using a combination of humans and machines," Mohan said. As an example, he explained that because glioblastoma tumors are highly heterogeneous and can infiltrate into adjacent, normal-appearing brain areas, human input is required to tell the model how and where to look for clues.
To validate the model, the researchers applied it to a cohort of 56 patients who had new lesions six months after beginning treatment. Patients were categorized as having pseudoprogression or tumor progression based on samples from repeat surgery, or if repeat tissue specimens were not available, two follow-up MRI scans using standard Response Assessment in Neuro-Oncology criteria. Patients then underwent diffusion tensor and dynamic susceptibility contrast-perfusion weighted MRI, and the researchers analyzed the results using the model, which distinguished the two groups of patients with 75.7 percent accuracy. The model's accuracy was more than 95 percent for classifying pseudoprogression.
Mohan believes the tool has potential to make a difference in glioblastoma survival rates for the first time in nearly 20 years. If doctors can diagnose early recurrence with confidence and intervene early, "the chances of changing the eventual outcome or improving overall survival for these patients" are greater, he said.
At Penn Medicine, doctors are now using the model to generate a score for glioblastoma patients receiving follow-up MRI scans to predict recurrence or tumor response. They then present the information to the institutional tumor board to make real-time clinical decisions for the patients.
As yet, Mohan said the tool is not ready for routine use in all glioblastoma patients, since it is not FDA approved and can't be billed to the patient's insurance. Instead, it is currently being used only for research.
Mohan's team is working with the FDA process to create a commercial version of the test. However, their goal is more expansive than simply detecting progression in glioblastoma patients. Instead, they're aiming to diagnose a total of 122 brain conditions. They also believe the model will have utility for tracking brain tumor progression with many types of treatment, including immunotherapy and other second-line therapies.
"We are doing 85 percent more MRI scans than we used to do 10 years ago, and there is not enough manpower to do that," Mohan said. "The vision we have for radiology is that radiologists will end up in more supervisory roles in the future, rather than looking at every image. Most of this work will be done by machines."