Machine learning approach allows for tumour classification during neurosurgery

Neurosurgical resection is the most common treatment for Central Nervous System tumours. However, there's a delicate balance between maximising the extent of the resection and minimising the neurosurgical damage, which depends on the tumour subtype. Currently, the neurosurgeon relies on preoperative imaging and the pathologist's intraoperative histological assessment on small tumour sections. However, this practice doesn't always result in a precise diagnosis, and post-operative revision might reveal that a second surgery is necessary or that, in hindsight, the resection could have been less radical.

A recent Nature publication reports on a new machine-learning-based approach for the intraoperative molecular subclassification of Central Nervous System tumours. The procedure starts by taking a small tumour tissue sample, followed by DNA extraction and nanopore sequencing. The sequencing procedure provides a sparse methylation profile, which is then fed into Sturgeon, a neural network classifier developed by the study's authors.

The performance of Sturgeon was assessed retrospectively on real nanopore sequencing data, resulting in the correct classification of 24 out of 27 paediatric tumours. The tests then moved into the operative theatre with a real-time intraoperative assessment during 25 surgeries performed at the Amsterdam University Medical Centers (AUMC) in the Netherlands. Sturgeon provided a correct diagnosis during 18 surgeries (72%) with a turnaround of 90 minutes. In the other cases, the model didn't reach the required confidence threshold.

This approach is fully compatible with the surgical timeline and is based on commercially available technologies and limited computational resource requirements. The insights provided by the model can be weighted by the pathologists, allowing the neurosurgeon to make a better-informed decision during the surgical procedure.

Reference: Vermeulen, C., Pagès-Gallego, M., Kester, L. et al. Ultra-fast deep-learned CNS tumour classification during surgery. Nature 622, 842–849 (2023).