Deep learning helps classify tumors during surgery

A rapid classification method for central nervous system (CNS) tumors that combines high-speed sequencing and deep-learned AI models could enable molecular diagnosis in less than 90 minutes, according to research published in . Nature. This finding demonstrates the potential of obtaining a molecular diagnosis of tumors during surgery to aid surgical decision-making.

Primary treatment for CNS tumors involves surgical removal of the tumor, which balances maximizing removal of tumor tissue while minimizing the risk of nerve damage and other complications. Requires careful consideration to take. Current standard practice relies on preoperative imaging and intraoperative histological analysis, but these methods are not always conclusive and sometimes inaccurate. . Sequencing DNA to reveal methylation profiles can reveal information about a tumor’s origin and prognosis, but results usually take several days to arrive.

To obtain DNA methylation profiles quickly enough to provide a diagnosis during surgery, Jeroen de Ridder and colleagues at the Oncode Institute in Utrecht, the Netherlands, used a technique called nanopore sequencing. Although this method is fast, the data it generates covers far fewer gene sites compared to traditional sequencing techniques. To enable molecular classification of CNS tumors using such sparse data, researchers developed a neural network tool named “Sturgeon.” “We developed Sturgeon, a patient-independent transfer learning neural network, to enable molecular subclassification of central nervous system tumors based on such sparse profiles.” is writing.

After training and validating the tool using simulated data, the authors tested Sturgeon on data from CNS tumor samples. Sturgeon correctly classified 45 out of 50 samples based on the equivalent of 20 to 40 minutes of sequencing. The authors also demonstrated the performance and applicability of Sturgeon in real time during 25 surgeries, achieving diagnostic turnaround times of less than 90 minutes. “Of these, 18 (72%) diagnoses were correct, but seven did not reach the required confidence threshold. We believe that machine learning diagnosis based on low-cost intraoperative sequences is We conclude that it has the potential to support neurosurgical decision-making, prevent neurological comorbidities, and avoid additional surgeries.”

The results of this study demonstrate that deep learning diagnostics based on fast intraoperative sequences may support neurosurgical decision-making and improve patient outcomes.