Neural Networks in Neurosurgery: The Role of AI as a Surgical Assistant

Reading time: 5 minutes

Anthony Tao

The modern computer has become a powerful part of society’s infrastructure. Among millions of other things, computers are critical for facilitating financial services, research, communications, and politics. However, we can all perhaps agree that one of their LEAST important functions is the ability to recognize cats. However, in 2012, the research team Google Brain trained their computer to accomplish just that ‒ to determine whether a picture contains a cat. Believe it or not, this represented one of the more exciting leaps in the field of artificial intelligence (A.I.), specifically in computer vision, the subfield of A.I. geared toward image processing.

In medicine, especially oncology, the development of reliable computer vision models has been most helpful in the realm of pathology. Given a biopsy of tumor tissue, A.I. can quickly recognize the severity and aggressiveness of a cancer, which will critically influence clinical decision-making. However, comprehensive cancer care also involves many other specialists such as radiologists, psychologists, and surgeons. The role of A.I. in these fields, especially surgery, is not as well developed.

Central nervous system (CNS) tumors ‒ that is, tumors of the brain or spinal cord ‒ typically require surgical removal. The extent and quality of the resection differs depending on the tumor type. For instance, CNS tumors harboring a defect in a gene called IDH oftentimes require a more complete resection of the tumor compared to others. This is because without normal IDH activity, a malicious molecule accumulates that enhances tumor formation. Unfortunately, there is no easy way of determining whether a given tumor contains the IDH mutation without first surgically opening the brain and retrieving a piece of tissue for biopsy. Therein lies the Catch-22: tumor type governs the surgical plan, but surgery is needed to determine tumor type.

The simple solution to such a problem would be to characterize the tumor tissue quickly during the surgery, which can last 2 to 5 hours. Once the tumor type is known, the surgeon can rapidly determine how the resection should proceed, altering the approach in real-time if needed. Unfortunately, it takes days for a tissue biopsy to be analyzed by a pathologist, and for the ultimate diagnosis to be returned to the surgeon.

To solve this problem, a group of researchers at UMC Utrecht in the Netherlands turned to A.I. and a type of technology called nanopore sequencing.

Like words in a book, DNA is composed of building blocks that instruct a cell’s functions. Interestingly, some of these “building blocks” can undergo a molecular modification called methylation. The specific DNA regions that are methylated can often define a cell’s identity ‒ for instance, whether the cell is a muscle cell, skin cell, or neuron. In cancer, this methylation pattern can also be used to identify tumor type. Nanopore sequencing, which uses nanoscopic sensors to “read” individual strands of DNA, is a tool that can quickly determine this methylation pattern.

Unfortunately, another problem arises. Though nanopore sequencing is fast, acquiring reliable methylation information still requires more time than a 2-hour surgery. Of course, a nanopore sequencing run can be cut short, but the resulting methylation information will be sparse. Thus, it would be useful if a tool could analyze this sparse information and still identify the tumor type with confidence.

This is where the researchers applied A.I., specifically a type of A.I. model known as a neural network, similar to the A.I. that drives Google Brain’s cat identifier as well as ChatGPT. They wanted to know if a neural network can be trained to interpret incomplete methylation information to accurately characterize a tumor sample ‒ and to do so quickly within the timeframe of a neurosurgical tumor resection.

The process of training a neural network requires a vast amount of training data. In the Google Brain initiative, researchers collected millions of images, manually classifying them as either images containing cats or those without cats. With a similar approach in mind, the UMC researchers took a bunch of sparse methylation data from cancer samples that have already been diagnosed. To learn how to predict tumor type using this information, the A.I. sifted through this data like flash cards; the front side represents a sample’s methylation pattern and the backside is the specific tumor type.

Once the A.I. was trained, the next step was to evaluate its performance. The researchers found that the A.I. was able to correctly label the tumor type of 80 out of 91 samples using only their sparse methylation profiles with very high confidence. They also found that the A.I. can theoretically perform well on methylation information from 15-20 minutes-worth of nanopore sequencing data. To directly assess the potential of this A.I. in real scenarios, they conducted the entire procedure ‒ from surgical tumor collection, to sample preparation, to sequencing, and ultimately to tumor classification with the A.I. ‒ for 25 CNS tumor surgeries (Figure 1). The average turnaround time for this entire procedure was 80 minutes, well within the timeframe of a 2-5 hour surgery.

This tool represents a powerful role with which A.I. can assist a surgeon during an operation. Moreover, the increasing efficiency with which nanopore sequencing can be accomplished can only further refine this application of A.I. Overall, as research and technology continue to advance, the collaboration between human expertise and A.I. promises a future where neurosurgery becomes more precise and informed, ushering in a new era of hope for patients battling brain tumors.

Edited by Charlotte Boyd

Work Discussed

Vermeulen, C. et al. Ultra-fast deep-learned CNS tumour classification during surgery. Nature 622, 842–849 (2023).