P13.15.B ULTRA-FAST DEEP-LEARNED CNS TUMOR CLASSIFICATION DURING SURGERY

Abstract

Abstract BACKGROUND Central nervous system (CNS) tumors are the most lethal category of tumors, particularly among children. Most commonly, the first line treatment of CNS tumors is neurosurgical resection of the tumor. During this procedure a delicate balance must be struck between extent of resection and risk of comorbidity. Ideally, the surgical plan is based on a detailed classification of the tumor. However this classification is often unknown at the start of the surgery. Current practice consists of preoperative imaging and intraoperative diagnosis achieved by rapid histological assessment of frozen tumor sections. However, these tests do not always result in a clear diagnosis, and are sometimes even revised based on postoperative tissue-based diagnostics. As a result, some patients require a second surgery, while others could in hindsight have been operated less radically. MATERIAL AND METHODS Nanopore sequencing enables direct detection of DNA methylation status in real time, making it ideally suited for intraoperative sequencing and molecular tumor classification. The major challenge of this approach is that the classification must be performed on very few randomly distributed sequence reads. To this end we developed Sturgeon, a neural-network training approach that delivers classification models capable of classifying 81 different CNS tumor subtypes based on sparse methylation data. Sturgeon leverages illumina 450K DNA methylation profiles to simulate millions of nanopore sequencing experiments, which are then used to train and validate classifier models. The resulting models are portable, work across patients and sequence depths and require less than a minute of computation on a laptop CPU. RESULTS We demonstrate our method to be fast enough to provide a correct diagnosis with as little as 20 to 40 minutes of sequencing in 45 out of 50 retrospective pediatric brain tumor samples. We applied Sturgeon in an intraoperative setting in 25 cases, where we consistently achieve a turnaround time of 60-90 minutes from biopsy to result with a correct molecular subclassification in 18 cases and an inconclusive result in 7 cases. CONCLUSION Sturgeon classification can be used in parallel with intraoperative frozen section histology. In cases where it agrees with the pathologist, Sturgeon offers a valuable independent confirmation, and a more detailed molecular subclassification. In the rare cases where Sturgeon disagrees with the histology diagnosis, or where the histology is inconclusive, it has the potential to prevent misdiagnoses that would result in unnecessary risk of comorbidity, or the requirement for a second surgery.