A groundbreaking study recently published in the journal Nature showcases the remarkable potential of artificial intelligence (AI) neural networks in revolutionizing the field of neurosurgery. The study, conducted by a team of researchers, introduces a patient-agnostic transfer-learned neural network known as Sturgeon. This innovative AI system enables ultra-fast molecular subclassification of central nervous system (CNS) tumors during surgery, providing neurosurgeons with crucial real-time information for more accurate decision-making.
CNS tumors require precise surgical resection, and current methods like preoperative imaging and intraoperative histological analysis may not always yield accurate results. In order to overcome these limitations, the researchers employed rapid nanopore sequencing to obtain sparse methylation profiles during surgery. However, the classification of these tumors remained challenging due to inadequate data and reference samples.
To tackle this issue, the researchers developed the Sturgeon machine learning classifier specifically designed for categorizing pediatric and adult CNS tumors. The team utilized nanopore sequencing data to train the neural network, dividing the dataset into different folds for submodel training, validation, and score calibration. The dataset consisted of 2,801 labeled methylation profiles from CNS tumor and normal tissue samples.
The Sturgeon neural network was trained using a curriculum learning method, starting with simpler simulations and gradually progressing to more complex ones. The network was fine-tuned using simulations with varying sparsity ranges, and it underwent 3,000 epochs of training and validation to ensure its accuracy. During inference, samples were categorized using four submodels, and the highest confidence level submodel’s scores were used for final classification.
The study yielded impressive results, with Sturgeon providing correct diagnoses for 45 out of 50 retrospectively sequenced samples within just 40 minutes of initiating sequencing. Moreover, the AI system demonstrated real-time effectiveness during 25 procedures, with a diagnostic turnaround time of less than 90 minutes. Out of these 25 procedures, 18 were accurately classified with a high confidence level.
It is worth noting that Sturgeon’s performance is directly linked to the depth of sequencing, and the neural network covers a significant portion of the genomic regions associated with CNS tumors within the first 50 minutes of sequencing simulations. The researchers also implemented temperature scaling to improve the overall calibration of the model.
The study showcases Sturgeon as the first AI model to successfully perform computationally intensive tasks such as training, validation, and calibration outside the surgical timeframe. This achievement has resulted in a highly accurate and well-tested one-size-fits-all classification model. The potential applications of this classifier extend beyond neurosurgery, offering rapid post-operative diagnostic capabilities and reducing turnaround times. Sturgeon’s usage in peripheral and low-income institutions holds great promise, although the required amount of tissue may prove to be a limiting factor.
In conclusion, the study highlights the immense potential of machine-learned diagnosis based on low-cost intraoperative sequencing in aiding neurosurgeons in decision-making. The Sturgeon neural network showcases its ability to accurately detect and classify tumor types within a short timeframe. Coupled with histological evaluation, this AI system can provide a more precise intraoperative diagnosis, particularly in challenging situations where the histology diagnosis is uncertain. The findings of this study open up new avenues for improving surgical outcomes, reducing comorbidity, and potentially avoiding unnecessary future procedures.
