2023-10-26 10:12:09
A team of researchers from University Medical Centers Amsterdam, Princess Máxima Pediatric Oncology Center and UMC Utrecht have developed a neural network that accelerates the identification of brain tumors, allowing neurosurgeons to adjust their strategies surgical. The study was recently published in Nature.
Cancer of the central nervous system (CNS) is one of the most serious cancers, especially in children. If neurosurgical resection of the tumor saves the lives of these patients, during the intervention, neurosurgeons do not know precisely what type of brain tumor it is or how aggressive it is. The exact diagnosis is usually only available one week following surgery, following the tumor tissue has been visually and molecularly analyzed by the pathologist.
Researchers used transfer learning to train Sturgeon, a patient-independent neural network, which enables molecular subclassification of CNS malignancies. They demonstrated its applicability in real time during 25 surgical procedures, obtaining a diagnosis in less than 90 minutes.
For the Princess Máxima Center, the results of the technique are sufficiently reliable and they are already using it. UMC Amsterdam will also use the technique in daily practice, to help speed up diagnosis.
Professor Eelco Hoving, pediatric neurosurgeon and clinical director of neuro-oncology at the Máxima Center, explains:
“During surgery, a small remnant of tumor tissue is sometimes deliberately left behind to prevent neurological damage. But if it later turns out, for example, that the tumor is very aggressive, a second surgery may still be necessary to remove this last remnant. This will once once more create risks and stress for patients and their families. This can be avoided now, because we will already know during the first intervention what type of tumor we are dealing with.”
Ultra-rapid classification of CNS tumors during surgery
Current approaches, such as preoperative imaging and intraoperative histological analysis, obtained by rapid histological evaluation of frozen tumor sections, do not prove to be completely reliable.
The researchers used Nanopore sequencing technology that reads DNA in real time and trained a system of four neural networks to analyze in real time and classify these sequences.
To overcome the limited number of data and reference samples, they created a labeled dataset of nearly 17 million artificial DNA sequences from 94 tumor types.
The models were trained on 36.8 million simulated nanopore cycles and validated on 4.2 million simulated nanopore cycles. The resulting Sturgeon model takes just seconds to run on a laptop.
The results
To comprehensively evaluate Sturgeon’s robustness and sensitivity to batch and operator bias, the researchers evaluated it on a publicly available nanopore sequencing dataset. Tested on samples taken during 25 procedures, it correctly classified 18 tumors, failing to reach the acceptable level of confidence for the other 7.
The algorithm might identify the type of tumor in 20 to 40 minutes. Researchers plan to retrain or refine it as new data becomes available, so it might identify other types of tumors. In the meantime, their approach has the potential to save many lives and contribute to patients’ quality of life.
Article references:
“Ultra-fast deep learning CNS tumor classification during surgery” Nature https://doi.org/10.1038/s41586-023-06615-2
Authors and affiliations:
C. Vermeulen, M. Pagès-Gallego, J. de Ridder, Oncode Institute, Utrecht, Center for molecular medicine, UMC Utrecht
L. Kester, MEG Kranendonk, J. van der Lugt, K. van Baarsen, EW Hoving, BBJ Tops, Center d’oncologie pédiatrique Princess Máxima, Utrecht, Pays-Bas
N. Verburg, P. de Witt Hamer, Department of Neurosurgery, University Medical Centers Amsterdam/VUmc
EJ Kooi, Department of Pathology, University Medical Centers Amsterdam/VUmc,
L. Dankmeijer, Department of Pathology, University Medical Centers Amsterdam/VUmc, Amsterdam, Department of Neurosurgery, University Medical Centers Amsterdam/VUmc, Amsterdam.
1698316589
#Researchers #deep #learning #classify #brain #tumors
