AI Tool Enables In-Surgery Genomic Profiling of Brain Tumor for Real-Time Guidance

Over-removal in the case of less aggressive tumors can negatively impact a patient's neurological and cognitive functioning, while under-removal in the case of highly aggressive ones can leave malignant tissue behind, resulting in rapid growth and spread. Current intraoperative diagnostic methods involve brain tissue freezing and microscopic examination, but these techniques often distort cell appearance and compromise clinical assessment accuracy. Furthermore, the human eye, even with advanced microscopes, can fail to reliably identify subtle genomic variations on a slide. Now, a novel artificial intelligence (AI) approach addresses these issues.

Scientists at Harvard Medical School (Boston, MA, USA) have developed an AI tool capable of swiftly decoding a brain tumor's DNA to determine its molecular identity during surgery. This process can take several days or even weeks using traditional methods. Having immediate access to a tumor's molecular type helps neurosurgeons decide on the extent of brain tissue removal and the application of tumor-killing drugs directly into the brain, all while the patient is still on the operating table. Modern advances in genomics have enabled pathologists to distinguish molecular signatures and associated behaviors among various brain cancer types. Aggressive glioma, for instance, has three main subvariants, each bearing unique molecular markers and growth propensities. Although AI models have been developed to profile other cancer types (e.g., colon, lung, breast), gliomas present unique challenges due to their molecular complexity and vast variation in tumor cell morphology.

The newly developed tool, named CHARM (Cryosection Histopathology Assessment and Review Machine), significantly expedites molecular diagnostics, which can be particularly useful in regions with limited access to technology for quick cancer genetic sequencing. CHARM was developed using 2,334 brain tumor samples from 1,524 individuals with glioma from three distinct patient populations. The tool exhibited a 93% accuracy rate when identifying tumors with specific molecular mutations in an unseen set of brain samples, and it successfully classified three major types of gliomas with distinct molecular features. Moreover, the tool was adept at visually analyzing tissue surrounding malignant cells, identifying areas of greater cellular density and higher cell death rates, both of which are indicators of more aggressive glioma types.

Additionally, CHARM was able to detect clinically important changes in a subset of low-grade gliomas, a less aggressive glioma subtype that is less likely to invade surrounding tissue. The tool further linked cellular appearance with the molecular profile of the tumor, thereby enabling the algorithm to determine how a cell's appearance relates to the tumor's molecular type. This comprehensive assessment improves the model's accuracy and mirrors how a human pathologist would visually evaluate a tumor sample. While CHARM was initially trained and tested on glioma samples, the researchers believe it can be successfully retrained to identify other brain cancer subtypes. However, the tool would require periodic retraining to reflect new disease classifications as they emerge from new findings. Although CHARM is freely available to other researchers, it needs clinical validation through real-world testing and FDA clearance before it can be used in hospitals.

“Right now, even state-of-the-art clinical practice cannot profile tumors molecularly during surgery. Our tool overcomes this challenge by extracting thus-far untapped biomedical signals from frozen pathology slides,” said study senior author Kun-Hsing Yu, assistant professor of biomedical informatics in the Blavatnik Institute at HMS.

Related Links:
Harvard Medical School