Histopathology-correlated evaluation of magnetic resonance spectroscopy for high-grade glioma prediction: A small-scale neuro-oncology study with exploratory machine learning.

Background: High-grade gliomas are aggressive primary brain tumours that require timely and accurate preoperative assessment. Magnetic resonance spectroscopy provides non-invasive metabolic information that may support neuro-oncological lesion characterisation. However, its tendency to show good sensitivity but limited specificity, particularly in the presence of metastasis, abscess, and other intracranial mimics, is already recognised. The clinical value of small, fully histopathology-confirmed cohorts lies in clarifying how these limitations appear in real diagnostic practice. Objective: This study evaluated the diagnostic performance of magnetic resonance spectroscopy for predicting high-grade glioma in a fully histopathology-confirmed cohort of patients with intracranial lesions. The study also aimed to characterise the false-positive pattern in clinically interpretable terms and to assess, as a secondary exploratory analysis, whether machine learning (ML) could add discriminatory value beyond the radiologist-derived magnetic resonance spectroscopy diagnosis. Methods: A retrospective cohort of 53 patients was analysed. All cases had histopathological confirmation, which was used as the reference standard. Patient-level variables included age, sex, magnetic resonance spectroscopy diagnosis, histopathological diagnosis, Cho/NAA, Cho/Cr, NAA/Cr, lactate, lipid, lesion location, enhancement pattern, necrosis, and oedema. Magnetic resonance spectroscopy was treated as the index test. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, F1-score, and balanced accuracy were calculated. Exploratory ML models were assessed using leave-one-out cross-validation. Results: The cohort included 31 males and 22 females, with a mean age of 48.1 ± 15.8 years. Histopathology confirmed high-grade glioma in 40 of 53 patients. Magnetic resonance spectroscopy classified 42 cases as high-grade glioma and 11 as non-high-grade glioma. Compared with histopathology, magnetic resonance spectroscopy achieved a sensitivity of 87.5%, specificity of 46.2%, positive predictive value of 83.3%, negative predictive value of 54.5%, accuracy of 77.4%, F1-score of 85.4%, and balanced accuracy of 66.8%. The seven false-positive cases were characterised as metastasis in three cases, abscess in three cases, and other non-high-grade pathology in one case. The best exploratory ML model was Random Forest with magnetic resonance spectroscopy diagnosis included, achieving an accuracy of 73.6% and an AUC of 0.646. However, this did not improve upon the diagnostic performance of radiologist-derived magnetic resonance spectroscopy diagnosis alone. Conclusion: In this small fully histopathology-confirmed neuro-oncology cohort, magnetic resonance spectroscopy demonstrated good sensitivity but limited specificity for predicting high-grade glioma. Rather than presenting this pattern as a new diagnostic phenomenon, this study contributes a clinically grounded confirmation with complete tissue verification, a clearly characterised false-positive mimic profile, and a transparent negative ML result. These findings support the use of magnetic resonance spectroscopy as an adjunctive metabolic neuroimaging tool and provide a cautious data point for future larger studies evaluating spectroscopy-guided and ML-supported diagnostic pathways.