Once their scalpels reach the edge of a brain tumor, surgeons face an agonizing decision: cut away some of the healthy brain tissue to ensure the entire tumor is removed, or leave the healthy tissue aside and risk leave behind some of the threatening cells. .
Now, scientists in the Netherlands report that they are using artificial intelligence to equip surgeons with knowledge about the tumor that can help them make that decision.
The method, described in a study published Wednesday in the journal NatureIt involves computer scanning segments of a tumor’s DNA and detecting certain chemical modifications that can produce a detailed diagnosis of the type and even subtype of the brain tumor.
That diagnosis, generated during the early stages of an hour-long surgery, can help surgeons decide how aggressively to operate, the researchers said. In the future, the method may also help guide doctors toward treatments tailored to a specific tumor subtype.
“It is imperative that the tumor subtype is known at the time of surgery,” said Jeroen de Ridder, an associate professor at the Center for Molecular Medicine at UMC Utrecht, a Dutch hospital, who helped lead the study. “What we have now uniquely achieved is to allow this very fine, robust and detailed diagnosis to be made already during surgery.”
Their deep learning system, called Sturgeon, was first tested on frozen tumor samples from previous brain cancer operations. He accurately diagnosed 45 of 50 cases within 40 minutes of starting genetic sequencing. In the other five cases she refrained from offering a diagnosis because the information was unclear.
The system was then tested during 25 live brain surgeries, most of them on children, along with the standard method of examining tumor samples under a microscope. The new approach yielded 18 correct diagnoses and failed to reach the necessary confidence threshold in the other seven cases. It resolved its diagnoses in less than 90 minutes, the study reported, short enough to inform decisions during an operation.
Today, in addition to examining brain tumor samples under a microscope, doctors can send them for more extensive genetic sequencing.
But not all hospitals have access to that technology. And even those who do, it can take several weeks before receiving results, said Dr. Alan Cohen, director of the Division of Pediatric Neurosurgery at Johns Hopkins and a cancer specialist.
“We have to start treatment without knowing what we are treating,” Dr. Cohen said.
The new method uses a faster genetic sequencing technique and applies it to only a small portion of the cell’s genome, allowing it to yield results before a surgeon has begun operating on the edges of a tumor.
Dr. de Ridder said the model was powerful enough to offer a diagnosis with sparse genetic data, similar to someone recognizing an image based on just one percent of its pixels and from an unknown portion of the image.
“You can figure out for yourself what you’re looking at and make a solid classification,” said Dr. de Ridder, who is also a principal investigator at the Oncode Institute, a cancer research center in the Netherlands.
But some tumors remain difficult to diagnose. Samples taken during surgery are about the size of a kernel of corn, and if they include some healthy brain tissue, the deep learning system may have difficulty selecting enough tumor-specific markers.
In the study, doctors addressed this by asking pathologists examining samples under a microscope to mark those with the most tumors for sequencing, said Marc Pagès-Gallego, a bioinformatician at UMC Utrecht and co-author of the study.
There may also be differences within a single patient’s tumor cells, meaning that the small segment that is sequenced may not be representative of the entire tumor. Some less common tumors may not correspond to those that have been previously classified. And some types of tumors are easier to classify than others.
Other medical centers have already begun applying the new method to surgical specimens, the study authors said, suggesting it may work in other people’s hands.
But Dr. Sebastian Brandner, professor of neuropathology at University College London, said sequencing and sorting tumor cells often still requires significant bioinformatics expertise, as well as workers who are able to execute, troubleshoot and repair the technology.
“The implementation itself is less straightforward than is often suggested,” he said.
Brain tumors are also the most suitable to be classified according to the chemical modifications analyzed by the new method; Not all cancers can be diagnosed that way.
The new method is part of a broad movement toward molecular precision in tumor diagnosis, which could allow scientists to develop targeted treatments that are less harmful to the nervous system. But translating deeper knowledge about tumors into new therapies has proven difficult.
“We’ve made some progress,” Dr. Cohen said, “but not as much in treatment as in understanding the molecular profile of tumors.”
