信息工程与人工智能学院知识库

肺癌是全球癌症死亡的主要原因，存活率极低，其中非小细胞肺癌（NSCLC）是肺癌的最大亚群，约占病例总数的85%且总生存率很极低，5年生存率仅为24%。近年来，随着PET/CT成像技术日益成熟，高通量测序技术的发展，获得基因表达谱变得较为方便，促进了基因特征的鉴定，这也为NSCLC的二级预防提供了充足准备。本实验共纳入160例NSCLC患者的放射基因组学数据，包括PET/CT图像、临床信息和RNA-seq数据。患者按照有无RNA-seq数据被分为训练组（108例有RNA-seq数据可用，随访期间有39例死亡）和测试组（52例无RNA-seq数据可用，随访期间有15例死亡）。实验流程首先通过特征挑选从普通患者的5268条RNA序列中筛选出9条RNA序列，接着参考PET/CT容量预后指数（PVP）模型方法将9个RNA序列整合为一个一维变量，即RNA预后指数。其次，依据训练数据的PET/CT和肿瘤掩膜图像，提取放射组学特征，并使用F检验方法降低特征维数。基于RNA指数选择的放射组学特征构建支持向量回归模型（SVR，使用Scikit-learn python包）。利用训练好的SVR模型预测52例患者的RNA预后指标。最后采用Cox比例风险模型评估分析RNA合成指数的预后价值。检验的最终结果显示，在训练数据的多因素Cox比例风险模型中，RNA预后指标与患者总生存期显著相关（风险比（HR）=2.602，P<0.001）。在测试数据的单因素Cox比例风险模型中RNA预后指数预测值与患者总生存期HR达到7.155，P-value值小于0.05，并且多因素生存分析中HR=8.8038，P-value值小于0.05，此结果证明RNA合成指数也与患者生存时间显著相关。此外为了证明RNA合成指数在多模态融合实验中对NSCLC患者复发预测的潜力。本研究将预测得到的52例患者RNA合成指数作为特征纳入NSCLC复发实验研究。基于3D卷积网络模型进行二分类实验，结果证明，综合RNA合成指数，影像和临床三种模态数据，AUC可达到0.70，相较于非RNA合成指数的方法，AUC提高约10%。本研究证明RNA合成指数在生存预后和复发研究中，都具有显著性的医学意义。

Lung cancer is the main cause of cancer death worldwide, with extremely low survival rates. Non small cell lung cancer (NSCLC) is the largest subgroup of lung cancer, accounting for approximately 85% of the total number of cases, and the overall survival rate is extremely low, with a 5-year survival rate of only 24%. In recent years, with the increasingly mature PET/CT imaging technology and the development of high-throughput sequencing technology, it has become more convenient to obtain gene expression profiles, promoting the identification of gene characteristics, which also provides sufficient preparation for the secondary prevention of NSCLC.This experiment included radiogenomic data from 160 NSCLC patients, including PET/CT images, clinical information, and RNA seq data. Patients were divided into a training group (108 cases with RNA-seq data available and 39 deaths during follow-up) and a testing group (52 cases without RNA-seq data available and 15 deaths during follow-up) based on the presence or absence of RNA-seq data. Firstly, 9 RNA-seq were screened from 5268 RNA sequences of common patients through feature selection, and then integrated into a one-dimensional variable, namely RNA prognostic index, by referring to the PET/CT volume prognostic index (PVP) model. Secondly, according to the PET/CT and tumor mask images of the training data, the radiomic features were extracted, and the F-test method was used to reduce the feature dimension. A support vector regression model (SVR, using the Scikit-learn python package) was constructed based on the RNA index and selected radiomic features. A trained SVR model was used to predict RNA prognostic indicators in 52 patients. Finally, Cox proportional hazard model was used to evaluate the prognostic value of RNA prognostic index. Final results of the examination showed that RNA prognostic markers were significantly associated with overall survival in a multivariate Cox proportional hazard model of the training data (HR=2.602, P<0.001). In the univariate Cox proportional hazard model of the test data, the predicted value of the RNA prognotic index and the overall survival of patients reached 7.155, and the P-value was less than 0.05; in the multivariate survival analysis, the HR=8.8038, and the P-value was less than 0.05. The results showed that RNA synthesis index was also significantly correlated with survival time. In addition, to demonstrate the potential of RNA synthesis index in multimodal fusion experiments for predicting recurrence in NSCLC patients. This study included the predicted RNA synthesis index of 52 patients as a feature in the NSCLC recurrence experimental study. Based on a 3D convolutional network model, a binary classification experiment was conducted, and the results showed that by combining RNA synthesis index, imaging, and clinical modal data, the AUC can reach 0.70, which is about 10% higher than the non RNA synthesis index method. This study demonstrates the significant medical significance of RNA synthesis index in survival prognosis and recurrence studies.