Quantitative assessment of the generalizability of a brain tumor Raman spectroscopy machine learning model to various tumor types including astrocytoma and oligodendroglioma.

Abstract

Significance: Maximal safe resection of brain tumors can be performed by neurosurgeons through the use of accurate and practical guidance tools that provide real-time information during surgery. Current established adjuvant intraoperative technologies include neuronavigation guidance, intraoperative imaging (MRI and ultrasound), and 5-ALA for fluorescence-guided surgery.

Aim: We have developed intraoperative Raman spectroscopy as a real-time decision support system for neurosurgical guidance in brain tumors. Using a machine learning model, trained on data from a multicenter clinical study involving 67 patients, the device achieved diagnostic accuracies of 91% for glioblastoma, 97% for brain metastases, and 96% for meningiomas. Here, the aim is to assess the generalizability of a predictive model trained with data from this study to other types of brain tumors.

Approach: A method was developed to assess the generalizability of the model, quantifying performance for tumors including astrocytoma, oligodendroglioma and ependymoma, pediatric glioblastoma, and classification of glioblastoma data acquired in the presence of 5-ALA induced fluorescence. Statistical analyses were conducted to assess the impact of vibrational bands beyond contributors identified in our previous research.

Results: A machine learning brain tumor detection model showed a positive predictive value (PPV) of 70% for astrocytoma, 74% for oligodendroglioma, and 100% for ependymoma. Furthermore, the PPV was 100% in classifying spectra from a pediatric glioblastoma and 90% for detecting adult glioblastoma labeled with 5-ALA-induced fluorescence. Univariate statistical analyses applied to individual vibrational bands demonstrated that the inclusion of Raman biomarkers unexploited to date had the potential to improve detectability, setting the stage for future advances.

Conclusions: Developing predictive models relying on the inelastic scattering contrast from a wider pool of Raman bands may improve detection accuracy for astrocytoma and oligodendroglioma. To do so, larger tumor datasets and a higher Raman photon signal-to-noise ratio may be required.