High-Precision Intraoperative Diagnosis of Gliomas: Integrating Imaging and Intraoperative Flow Cytometry with Machine Learning

Shunichi Koriyama¹Yutaka Matsui²Takahiro Shioyama³Mikoto Onodera⁴Manabu Tamura⁵Tatsuya Kobayashi⁶Buntou Ro⁶Kenta Masui⁷Takashi Komori⁸Yoshihiro Muragaki^9*Takakazu Kawamata⁶

• ¹Tokyo Women's Medical University, Shinjuku, Japan
• ²Technical Department, Atom Medical Corporation, Tokyo, Japan
• ³Ogino Memorial Laboratory, Nihon Kohden Corporation, Tokyo, Japan
• ⁴Faculty of Advanced Techno-Surgery (FATS), Institute of Advanced Biomedical Engineering & Science, Graduate School of Medicine, Tokyo Women's Medical University, Tokyo, Japan
• ⁵Faculty of Advanced Techno-Surgery (FATS), Institute of Advanced Biomedical Engineering & Science, Graduate School of Medicine, Tokyo Women's Medical University,, Tokyo, Japan
• ⁶Department of Neurosurgery, Tokyo Women's Medical University, Tokyo, Japan
• ⁷Department of Integrated Neuroscience, Graduate School of Biomedical Sciences, Nagasaki University, Nagasaki, Japan
• ⁸Department of Laboratory Medicine and Pathology, Tokyo Metropolitan Neurological Hospital, Tokyo Metropolitan Hospital Organization, Tokyo, Japan
• ⁹Center for Advanced Medical Engineering Research and Development, Kobe University, Hyogo, Japan

Introduction: Accurate intraoperative identification of glioma molecular subtypes, such as isocitrate dehydrogenase mutation and 1p/19q co-deletion, is essential for precise diagnosis, prognostication, and determining the extent of tumor resection-balancing maximal tumor removal with preservation of neurological function. Methods: We developed a machine learning model that integrates preoperative imaging features (magnetic resonance imaging, computed tomography, and ^11Cmethionine positron emission tomography [PET]) and intraoperative flow cytometry (iFC) data to predict molecular subtypes of glioma in real-time. Results: Analyzing 288 cases of diffuse gliomas, this model achieved an overall accuracy of 76.0%, with a macro-average ROC-AUC of 0.88 and a micro-average ROC-AUC of 0.89. Key predictive factors included the tumor-to-normal uptake ratio on PET, malignancy index from iFC, and patient age, all of which showed significant differences between correctly and incorrectly classified cases. We also developed a prototype application that visualizes the prediction results intraoperatively, thereby supporting real-time surgical decisionmaking. Conclusion: This integrated approach enhances the precision of intraoperative molecular diagnosis and has the potential to optimize surgical strategies for glioma treatment.