NEW YORK – A multi-institution research team said that it has developed a tool that could be used during brain tumor removal surgeries to identify within seconds the extent of cancer infiltration into surrounding tissue and aid decisions on whether to continue resection.
The team led by scientists at the University of California San Francisco and the University of Michigan reported last week in Nature that they had used an artificial intelligence-developed algorithm to analyze stimulated Raman histology (SRH) slide images of unprocessed tissue specimens that had been taken from the margins of surgery sites during the removal of tumors from patients with diffuse glioma. They reported that the model, dubbed FastGlioma, was used to stratify samples by the extent of tumor infiltration, and the results outperformed some of the most commonly used adjuncts for the identification of high-risk tissue.
Shawn Hervey-Jumper, a senior author on the study and a neurosurgeon at UCSF, noted that gliomas become integrated with the surrounding healthy brain cells rather than developing well-defined borders. The issue is common among brain cancers as well as various other types of cancers, said co-senior author Todd Hollon, a neurosurgeon at the University of Michigan, and most patients who undergo tumor removal surgery will have residual tumor tissue that should have been removed from the resection cavity.
"We know that better surgeries result in better outcomes," he said. "So, patients who have more of the tumor removed will have longer overall survival, better quality of life, longer progression-free survival."
Hollon was also the lead author of a 2020 article that reported that SRH-based imaging with an AI-based algorithm could classify surgical samples from brain tumors as accurately as a pathologist. That article was published in Nature Medicine.
During the recent study, the authors used the portable SRH-based Invenio Nio instrument to generate slide images that are similar in appearance to hematoxylin and eosin slide images. SRH is a label-free optical imaging method that provides sub-micrometer resolution and generates contrast in the images through differences in biochemical properties within the specimen.
Hervey-Jumper said that FastGlioma's algorithm was used to analyze the image and determine whether segments on the slide likely showed tumor or healthy tissue with results in about 10 seconds. It analyzes microscopic features, such as hypercellularity and cytological features of the cells, and the model is designed to detect those features across various types of gliomas, Hollon added. The authors noted that they used high-speed, low-resolution instrument settings to reduce the time to results compared to high-resolution imaging as well as compared to conventional microscopic analysis.
Hollon said that the researchers' approach of using SRH-based imaging of unprocessed samples has the potential to give surgeons microscopic information without the 30 or 40 minutes of delays to prepare a frozen section. By decreasing the resolution of the captured images, they were able to further reduce the time from minutes to seconds for the analysis of each slide with minimal loss in accuracy, he said.
While interoperative MRI and fluorescence-guided surgery have been validated as surgical adjuncts for use in markets including the US and EU, Hollon said that those tools have limitations. He noted that contrast-enhanced MRI images can be noisy, indirect indicators of tumor infiltration and can have low sensitivity and specificity.
While the compound 5-aminolevulinic acid (5-ALA) is used as an adjunct for fluorescence-guided removal of malignant glioma cells, Hervey-Jumper noted that it only works with the most aggressive tumors, and the results don't include any measurements of tumor burden. No dyes are on the market to help identify what portion of tumors are removed from a patient with grade 2 or 3 gliomas, he said.
The prospective study involved the examination of slide samples from 220 adult patients with diffuse glioma who underwent tumor resection. The authors wrote that FastGlioma achieved a mean area under the receiver operating curve of 92.1 percent for the stratification of diffuse glioma infiltration by numerical scores from 0 for normal brain tissue to 3 for dense tumor infiltration, and the model performance and scoring was consistent across patient demographics.
"Importantly, visual foundation model pretraining allowed FastGlioma to generalize to the fast, low-resolution images acquired at 10X imaging speed without a clinically significant reduction in prediction performance," they said.
The researchers also found that FastGlioma was less likely to miss high-risk tissue compared with the most commonly used surgical adjuncts. In a simulated prospective surgical trial, the researchers used the tool to evaluate 624 samples from 129 patients and compared the results against the performance of image-guided surgery with MRI-based neuro-navigation and fluorescence-guided surgery with 5-ALA.
Following SRH imaging, the specimens were extracted and sent for further downstream whole-slide analysis using H&E and immunohistology testing, and specimen-level tumor-infiltration scores were determined to compare the results across the three surgical adjuncts.
The authors said that FastGlioma achieved an AUROC of 98.1 percent compared to 76.3 percent for positivity through MRI fluid-attenuated inversion recovery (FLAIR) results, 71.8 percent for MRI-based contrast enhancement, and 89.0 percent for 5-ALA fluorescence. The researchers also determined that five of the 129 patients, about 4 percent, who were tested with FastGlioma had at least one high-risk tumor that was left within the resection cavity following surgery compared to 31 of 129, or 24 percent, among the other adjuncts.
"Errors on this task are clinical high-risk errors because they represent actionable and decisive predictions: normal brain predictions signal to stop resection, dense tumor signals to continue resection if otherwise safe," they wrote.
The authors further noted that the interpretation of MRI FLAIR results can be difficult during diffuse glioma surgery because tumor infiltration or cerebral edema can result in positivity. FastGlioma, however, was used during the study to differentiate between tumor infiltration and cerebral edema in FLAIR-positive regions with an AUROC of 98.7 percent.
Hollon said the team pushed the limits of how quickly they could acquire a slide image and still deliver accurate predictions of tumor infiltration, and the algorithmic model had been resistant to the image degradation that was introduced with lowered resolution. He said that a surgeon would likely use the methods to analyze about 10 images for a surgery.
Hervey-Jumper said that the model was slightly more accurate for the detection of glioblastoma and other aggressive types of brain tumors compared with less aggressive gliomas that look more similar to healthy brain tissue.
Hollon said that he sees the potential to apply the same methods to other types of cancer surgeries that require careful delineation of tumor margins, especially head and neck tumor removal surgeries. He also said, however, that the same methods could be developed into a general tool to aid clinicians who are removing myriad types of cancer tissues.
Meanwhile, he is interested in further developing the glioma-specific application to determine whether it can provide prognostic results on progression-free and overall survival. Those measures could leverage the tumor infiltration scores as well as other predictive biomarkers in the slide images.
"If we know at the end of that operation that at one of the margins, you still have 40 percent tumor infiltration or 80 percent tumor infiltration, whatever it may be, we think that that is going to be relevant for predicting the risk that that patient has for treatment failure, … recurrence, and then, potentially, decreased survival," he said.
Hollon said that his team and the University of Michigan hope to get these methods into the hands of surgeons. He said that the researchers have not formed a company, but he thinks that a licensing deal could help bring the FastGlioma test to market.
Hervey-Jumper noted, though, that the tools need further study, and that surgeons will still be unable to remove areas where cancer has infiltrated important structures of the brain. "We'll still have recurrences locally but hopefully fewer of them in areas where we just couldn't see well enough," he said.