Researchers have introduced an AI-driven model capable of quickly determining, in just 10 seconds during surgery, whether any part of a cancerous brain tumor remains that could still be excised. This technology, named FastGlioma, has significantly outperformed traditional techniques used to assess leftover tumor tissue. Experts believe it has the potential to revolutionize neurosurgery by enhancing patient management for those with diffuse gliomas.
Researchers have unveiled an AI-based model that can ascertain in just 10 seconds during surgery if any removable parts of a cancerous brain tumor are left, as indicated by a study published in Nature.
The technology, known as FastGlioma, showed a significant advantage over traditional techniques for detecting remaining tumor tissue, according to a team led by experts from the University of Michigan and the University of California San Francisco.
Senior author Todd Hollon, M.D., a neurosurgeon at the University of Michigan Health and assistant professor at U-M Medical School, stated, “FastGlioma is an AI-based diagnostic tool that has the potential to transform neurosurgery by instantly improving the management of patients with diffuse gliomas.”
“This technology operates more quickly and precisely than the current standard methods for tumor detection and could be applied to diagnose other brain tumors in both children and adults. It could serve as a foundational tool for guiding brain tumor surgeries.”
When neurosurgeons remove dangerous tumors from patients’ brains, they rarely manage to eliminate the entire tumor mass.
The remnants are referred to as the residual tumor.
Often, the remaining tumor goes undetected during surgery because surgeons struggle to differentiate between healthy brain tissue and the residual tumor in the cavity that once housed the mass.
The residual tumor can closely resemble healthy brain tissue, presenting a significant challenge during surgery.
Neurosurgical teams utilize various techniques to identify residual tumors during procedures.
They may perform MRI scans, which require specialized equipment that is not universally available.
Another approach is the use of fluorescent imaging agents to visualize tumor tissue, but this method is not suitable for all tumor types.
These limitations hinder broader adoption of existing methods.
In an international study of this AI-based technology, neurosurgical teams examined fresh, unprocessed samples from 220 patients who underwent surgery for either low- or high-grade diffuse glioma.
FastGlioma was found to detect and quantify the remaining tumor with an average accuracy of around 92%.
In comparison with surgeries guided by predictions from FastGlioma versus traditional imaging and fluorescent techniques, the AI technology missed high-risk residual tumors only 3.8% of the time, whereas conventional methods had nearly a 25% miss rate.
Co-senior author Shawn Hervey-Jumper, M.D., a professor of neurosurgery at the University of California San Francisco, commented, “This model is a groundbreaking advancement over current surgical methods, rapidly identifying tumor infiltration at a microscopic level using AI and dramatically reducing the chances of overlooking residual tumor during glioma resection.”
“With FastGlioma, we can reduce reliance on conventional imaging technologies, contrast agents, or fluorescent markers to maximize tumor removal.”
How it works
FastGlioma assesses remaining tumor tissue by merging microscopic optical imaging with a type of AI called foundation models.
Foundation models, like GPT-4 and DALL·E 3, are advanced AI systems trained on extensive and diverse datasets that can adapt to various tasks.
After extensive training, these models can classify images, engage in conversations, respond to emails, and create images based on textual descriptions.
To create FastGlioma, researchers pre-trained the visual foundation model using over 11,000 surgical specimens and four million distinct microscopic images.
The tumor specimens are examined using stimulated Raman histology, a rapid and high-resolution imaging technique developed at U-M.
This same approach was employed to develop DeepGlioma, another AI diagnostic system that identifies genetic mutations in brain tumors in under 90 seconds.
Honglak Lee, Ph.D., co-author and professor of computer science and engineering at U-M, noted, “FastGlioma can identify residual tumor tissue without depending on lengthy histology protocols and large labeled medical datasets, which are often limited.”
Obtaining full-resolution images through stimulated Raman histology takes approximately 100 seconds, while a “fast mode” offers lower-resolution images in just 10 seconds.
Researchers reported that the full-resolution model achieved an accuracy of up to 92%, with the fast mode slightly lower at around 90%.
“This indicates that we can identify tumor infiltration within seconds with very high accuracy, which can inform surgeons whether further resection is necessary during an operation,” Hollon added.
The future of AI in cancer care
Over the past two decades, the occurrence of residual tumor post-neurosurgery has not markedly improved.
Residual tumors not only contribute to a diminished quality of life and earlier mortality for patients but also add stress to a healthcare system expecting 45 million surgical procedures worldwide annually by 2030.
Global cancer initiatives have suggested integrating emerging technologies, including advanced imaging techniques and AI, into cancer surgery.
The Lancet Oncology Commission on global cancer surgery highlighted in 2015 the urgent need for cost-effective strategies to address surgical margins in cancer procedures, emphasizing the importance of new technologies.
FastGlioma not only serves as an accessible and affordable tool for neurosurgical teams dealing with gliomas, but researchers also find it capable of effectively detecting residual tumor in various non-glioma cancers, such as pediatric brain tumors like medulloblastoma and ependymoma, as well as meningiomas.
Co-author Aditya S. Pandey, M.D., chair of the Department of Neurosurgery at U-M Health, remarked, “These findings underscore the potential of visual foundation models like FastGlioma for medical AI applications, allowing for broader application to other human cancers without needing extensive retraining or adjustments.”
“Future research will focus on utilizing the FastGlioma framework for other cancers, including lung, prostate, breast, and head and neck cancers.”
