Ultra-fast deep-learned CNS tumour classification during surgery

Area of Science:

• Neuro-oncology
• Artificial Intelligence in Medicine
• Genomic Medicine

Background:

• Central nervous system (CNS) tumors are highly lethal, especially in children.
• Neurosurgical tumor resection requires balancing maximal tumor removal with minimal neurological damage.
• Current diagnostic methods like imaging and histology are often inconclusive or inaccurate.

Purpose of the Study:

• To develop a patient-agnostic, transfer-learned neural network (Sturgeon) for molecular subclassification of CNS tumors.
• To enable rapid, intraoperative CNS tumor diagnosis using sparse methylation profiles from nanopore sequencing.
• To assess the accuracy and speed of Sturgeon in real-time surgical settings.

Main Methods:

• Development of Sturgeon, a neural network trained on sparse methylation profiles.
• Utilizing rapid nanopore sequencing for intraoperative data acquisition.
• Validation of Sturgeon on retrospective and real-time surgical cases.

Main Results:

• Sturgeon achieved accurate diagnoses in 45/50 retrospective samples within 40 minutes.
• In real-time surgeries, Sturgeon provided diagnoses in under 90 minutes.
• 72% (18/25) of real-time diagnoses were correct, with others not reaching confidence thresholds.

Conclusions:

• Machine-learned diagnosis via low-cost intraoperative sequencing can aid neurosurgical decision-making.
• This approach has the potential to prevent neurological comorbidity and avoid repeat surgeries.
• AI-powered molecular subclassification offers a promising advancement in CNS tumor management.