Machine learning–assisted intraoperative hyperspectral imaging in cranial meningioma surgery: interim results of an ongoing study

OBJECTIVE Hyperspectral imaging (HSI) is a label-free optical technique that might extend conventional intraoperative visualization by providing quantitative spectral information from the operative field. The aim of this study was to evaluate whether intraoperative in vivo HSI combined with machine learning (ML)–based analysis can distinguish cranial meningioma from adjacent intracranial tissues, particularly cortex and dura, during meningioma surgery. METHODS This interim analysis of an ongoing prospective observational study included patients who underwent elective cranial meningioma surgery at a single tertiary neurosurgical center. Intraoperative HSI was performed using a CE-marked tissue system at predefined surgical time points. Expert-defined regions of interest (ROIs) representing meningioma, cortex, and dura were annotated on HSI-derived RGB images. Pixel-wise reflectance spectra were extracted and analyzed as raw reflectance spectra, standard normal variate–normalized spectra, and first- and second-derivative spectra. Four ML classifiers, including linear discriminant analysis, support vector machine, logistic regression, and multilayer perceptron, were evaluated using leave-one-patient-out cross-validation. Two classification tasks were investigated: binary discrimination of meningioma versus cortex, and multiclass discrimination of meningioma, cortex, and dura. RESULTS Among 70 patients, 291,104 meningioma pixels from 248 ROIs, 141,515 cortex pixels from 135 ROIs, and 333,244 dura pixels from 104 ROIs were analyzed. Meningioma showed distinct spectral characteristics, including attenuation of the oxyhemoglobin-associated double peak and lower tissue oxygenation, compared with cortex and dura. For binary classification of meningioma versus cortex, the multilayer perceptron using standard normal variate–normalized spectra achieved the highest performance, with an accuracy of 0.928, precision of 0.960, recall of 0.932, F1 score of 0.946, weighted F1 score of 0.929, and Matthews correlation coefficient of 0.840. In the multiclass task, performance was lower but remained feasible, with the same classifier-preprocessing combination achieving an accuracy of 0.773, macro F1 score of 0.782, weighted F1 score of 0.772, and macro Matthews correlation coefficient of 0.662. CONCLUSIONS ML-assisted intraoperative HSI enabled feasible discrimination of expert-annotated meningioma, cortex, and dura regions in cranial meningioma surgery. These interim findings support HSI as a promising label-free optical guidance technology, while underscoring the need for external validation, patient-level analyses, WHO grade prediction, and spatially matched histopathological confirmation before clinical implementation.