【摘要】 脑胶质瘤是最常见的原发性中枢神经系统肿瘤，具有高度的侵袭性。其中胶质母细胞瘤(GBM)是脑胶质瘤中恶性程度最高的一种，患者在5年内存活率只有5.6%。表皮生长因子受体(EGFR)对脑胶质瘤的生长、侵袭和复发中起着重要作用，在胶质母细胞瘤中，EGFR扩增和突变已被确定为驱动因素。目前脑胶质瘤整合诊断流程受限于实验操作复杂，往往存在一定滞后性，需在手术后2周左右才能得到结果，无法为术者提供实时分子病理信息支持。本文提出了一种基于术中病理冰冻切片的太赫兹时域光谱(THz-TDS)数据结合卷积神经网络(CNN)对EGFR扩增状态进行预测的方法。术中通过THz-TDS系统采集脑胶质瘤冰冻切片的光谱数据，计算其吸收系数，并利用Savitzky-Golay滤波器将其平滑处理后，再用格拉姆角场(GAF)、马尔可夫转移场(MTF)和递归图(RP)将吸收系数分别转化成二维图像数据作为后续CNN模型的输入。为充分利用图像数据，我们采用单一图像输入、前端融合和中端融合等不同方式搭建CNN模型。通过对比分析不同模型下的受试者工作特征(ROC)曲线下面积(AUC)值发现，格拉姆角和场(GASF)与格拉姆角差场(GADF)的中端融合卷积神经网络模型预测效果最好，测试集预测的AUC值为94.74%。此外，目前常用的基于太赫兹光谱数据的预测模型中，多是利用一维光谱数据降维后结合机器学习进行分析，处理过程中会造成部分数据信息丢失。因此我们还对吸收系数结合机器学习的方法进行了训练和测试。通过对比一维数据和二维图像的不同模型结果，可以发现相较于一维太赫兹时域光谱数据进行机器学习，二维光谱图像在卷积神经网络中训练模型有着更好的预测效果。实验结果表明本文提出的基于太赫兹光谱数据和卷积神经网络模型能够实现EGFR扩增状态的实时快速预测，为研究太赫兹时域光谱在脑胶质瘤中进行分子病理学分类提供了新的思路，对术中及时调整手术策略以及尽早制定术后辅助治疗方案具有重要意义。更多还原

【Abstract】 Gliomas are the most common primary central nervous system tumors with high invasiveness. Glioblastoma(GBM) is the most malignant type of brain glioma, with a 5-year survival rate of only 5.6%. The epidermal growth factor receptor(EGFR) plays an important role in the growth, invasion, and recurrence of glioblastoma. EGFR amplification and mutation have been identified as driving factors in glioblastoma. Currently, the integrated diagnosis process for glioma is limited by complex experimental procedures, often with a certain lag, and results can only be obtained approximately 2 weeks after surgery, which does not provide real-time molecular pathological information support for the operator. This article proposes a method for predicting EGFR amplification status based on intraoperative pathological frozen sections using terahertz time-domain spectroscopy(THz-TDS) data combined with convolutional neural networks(CNN). During the operation, spectral data of frozen sections of brain gliomas were collected using the THz-TDS system, and their absorption coefficients were calculated. After smoothing using the Savitzky-Golay filter, the absorption coefficients were converted into two-dimensional image data using the Gram Angular Field(GAF), Markov Transition Field(MTF), and Recursive Plots(RP) as inputs for subsequent CNN models. To fully utilize image data, we employ various methods, including single-image input, front-end fusion, and mid-range fusion, to construct CNN models. By comparing and analyzing the Area Under the Curve(AUC) values of Receiver Operating Characteristic(ROC) curves under different models, it was found that the Mid range Fusion Convolutional Neural Network model with Gram Angular Summation Field(GASF) and Gram Angular Difference Field(GADF) had the best prediction performance, with a predicted AUC value of 94.74% in the test set. In addition, the commonly used prediction models based on terahertz spectral data often-employ one-dimensional spectral data for dimensionality reduction and machine learning analysis, which may result in partial loss of data information during processing. Therefore, we also trained and tested the method of combining the absorption coefficient with machine learning. By comparing the results of different models for one-dimensional data and two-dimensional images, it is found that training models with two-dimensional spectral images in convolutional neural networks yields better predictive performance compared to machine learning with one-dimensional terahertz time-domain spectral data. The experimental results-demonstrate that the proposed method, based on terahertz spectroscopy data and a convolutional neural network model, can achieve real-time and rapid prediction of EGFR amplification status, providing new insights for molecular pathological classification of brain gliomas using terahertz time-domain spectroscopy. It is of great significance for the timely adjustment of surgical strategies during surgery and the early development of postoperative adjuvant treatment plans.更多还原