【摘要】 脑胶质瘤是原发于中枢神经系统的恶性肿瘤，其恶性进展快，致死致残率高，目前主要依赖外科手术联合术后放化疗，缺乏准确和灵敏的手段在术中快速判断脑胶质瘤的组织及分子病理类型以指导手术策略。采用透射式太赫兹时域光谱（THz-TDS）探测系统对术中病理样本进行快速测试，实验结果发现：在特定频段内正常脑组织、肿瘤和肿瘤周边组织对太赫兹波吸收率有较大差异，以此可以实现对肿瘤不同部位进行区分，更精准地确定手术范围；建立太赫兹光谱数据库，对肿瘤的分子分型进行快速预测，测试集的AUC值可达0.907，从而为手术方式及术后治疗快速提供依据。更多还原

【Abstract】 Objective Glioma is a highly malignant tumor originating from the central nervous system, characterized by rapid progression and high rates of mortality and disability. The standard treatment involves surgical resection followed by postoperative radiotherapy and chemotherapy. Intraoperative rapid frozen pathological diagnosis plays a critical role in guiding surgical strategy but suffers from limited sensitivity. Although histopathological and molecular pathological tests offer high accuracy and sensitivity, they typically require at least one week to yield results, thereby failing to provide real-time intraoperative guidance on tissue classification and molecular subtypes. This study aims to address this clinical gap by investigating the application of terahertz spectroscopy for real-time intraoperative pathological diagnosis and molecular typing of glioma, enabling accurate differentiation of tumor-related tissues and rapid prediction of key molecular markers to support intraoperative decision-making and inform postoperative therapeutic planning.Methods Fresh surgical specimens from 32 glioma patients(approved by the ethics committee with informed consent) are collected, immediately stored in pre-cooled PBS, frozen-embedded at-80 ℃, and sectioned into 100 μm slices for terahertz measurement(mounted on 2 mm PE sheets) and 10 μm slices for H&E staining. A transmission terahertz time-domain spectroscopy(THz-TDS) system is employed to acquire absorption spectra in the 0.2-1.4 THz range, with the optical path enclosed in a nitrogen-filled chamber(relative humidity <3%, temperature is 20 ℃±0.2 ℃) to minimize water vapor interference, and a custom six-hole sample holder is used to ensure efficient and stable measurements. Time-domain signals are converted to the frequency domain via Fourier transform to calculate absorption coefficients, refractive indices, and extinction coefficients. Statistical analysis begins with the KolmogorovSmirnov(K-S) test for normality, followed by either an independent samples t-test or a nonparametric rank-sum test, using the integrated absorption coefficient across the full spectral band as the test variable. Terahertz spectral data are transformed into Gramian angular summation field(GASF), Gramian angular difference field(GADF), Markov transition field(MTF), and recurrence plot(RP) images. Subsequently, single-input CNN, feature-fusion CNN(FFCNN), and multi-feature CNN(MFCNN) models are constructed to predict molecular markers such as IDH mutation status, with concurrent integration of pathological grading, histological subtypes, and seven key molecular pathological indicators(e.g., IDH mutation, MGMT promoter methylation) for correlation analysis.Results and Discussions The study identifies significant differences in terahertz absorption rates among normal brain tissue, peritumoral tissue, and tumor core tissue across the 0.2-1.4 THz band. Normal brain tissue exhibits absorption rates similar to those of peritumoral tissue across different WHO grades but significantly higher than those of tumor core tissue; however, the difference between normal brain tissue and WHO grade IV tumor core tissue is smaller than that observed with WHO grades I and III. Peritumoral tissue consistently demonstrates higher absorption rates than tumor core tissue across all pathological grades and types, with this disparity being more pronounced in WHO grade IV gliomas and between glioblastoma(GBM) and other pathological subtypes(e. g., astrocytoma, oligoastrocytoma) in peritumoral tissue than in tumor core tissue. Terahertz absorption rates increase with ascending WHO grade, and GBM exhibits higher absorption rates than other subtypes. Additionally, significant spectral differences are observed between tumor tissues stratified by molecular marker status(e. g., IDH mutation, MGMT methylation, 1p/19q co-deletion). The MFCNN model, constructed by fusing GASF and GADF images derived from spectral data, achieves the best performance in predicting IDH mutation status, with an AUC value of 0.907, outperforming both single-input CNN and FFCNN models. These findings underscore the potential of terahertz spectroscopy to indirectly capture microstructural and molecular-level alterations in tissue, with peritumoral tissue containing richer pathological and molecular information than tumor core tissue—an observation rarely reported systematically in prior studies, which are largely limited to animal models or paraffin-embedded specimens. The observed spectral differences are attributed to factors including tissue water content, edema, necrosis, and molecular composition changes driven by molecular markers, thereby providing a physical basis for intraoperative applications.Conclusions This study demonstrates that terahertz spectroscopy technology can effectively distinguish normal brain tissue, peritumoral tissue, and tumor core tissue in the 0.2-1.4 THz band, with greater spectral variability observed in peritumoral tissue that increases with glioma WHO grade. By integrating machine learning and deep learning frameworks, robust prediction models are successfully developed to enable rapid and accurate inference of key molecular markers such as TP53 and IDH, thereby validating the feasibility of terahertz spectroscopy for intraoperative real-time pathological diagnosis and molecular subtyping of glioma. This approach overcomes the limitations of conventional intraoperative diagnostic methods and provides a novel strategy to guide precise tumor resection and individualized therapeutic interventions during glioma surgery. Future work will focus on expanding the sample cohort, incorporating additional molecular markers, refining tissue processing and measurement protocols, and elucidating the underlying mechanisms that link molecular alterations to terahertz spectral features to facilitate clinical translation and broader application of this technology in intraoperative precision diagnosis and treatment of glioma.更多还原