基于放射组学对胰腺癌分子标记物表型及预后预测研究

Some specific molecular may play an important role in the occurrence, development and metastasis of pancreatic cancer (PC). Some molecular markers of PC such as K-ras、c-myc、C-erbB-2、P53、P16、nm23、SMAD4、EGFR、Ki-67 has been widely used in clinical diagnosis and therapy. Tissue diagnosis can be made using either histologic or cytological analysis. Histology requires a tissue specimen, and allows assessment of both architecture and cell morphology. Cytology is the analysis of individual cells that are exfoliated, aspirated from a fine needle, or obtained from peritoneal washings. But, these technologies are invasive, tissue sample limited, pancreatic fistula, the results greatly influenced by the operator and the location of tumor. Therefore, the clinical use is greatly limited. Now it is urgent to find a noninvasive, convenient, accurate and effective diagnostic method for achieving PC molecular information before the surgery. In our preliminary study, we have find that the model by radiomics showed a good discrimination to P53, yielding an accuracy of 92.5%.. This project will investigate the utility a radiomic analysis of hundreds of features quantifying tumor image intensity, shape and texture in CT volume scans. Then the optimal features will be selected. Finally the optimal features which can exhibit strong phenotype will be selected by neural net algorithm. This project aim to provide a quantitative radiomics approach to clinicians for decoding molecular markers phenotype, assessing the prognosis of PC and potentially the patient strategy in personalized medicine.

胰腺癌（pancreatic carcinoma，PC）的发生、发展和预后与某些特定分子标记物（K-ras、c-myc、C-erbB-2、P53、nm23、SMAD4、表皮生长因子受体、Ki-67等）密切相关。但是，术前获得此类分子标记物信息需要通过穿刺活检或胰液细胞学检查，它们具有侵入性、获取组织有限、并发症多、受操作者和肿瘤部位影响大，运用于临床受到极大限制。目前急需一种无创、准确的新方法在术前获取PC分子标记物信息。我们前期研究表明通过放射组学可以对PC分子标记物P53进行准确预测，准确率达92.5%。 基于此，本课题拟采用高分别率CT扫描，分割出肿瘤3D图像，高通量提取图像纹理参数，进行最优特征选择，通过建立人工神经网络模型成功实现对PC重要分子标记物和预后的预测。本项目可能为PC患者术前准确进行分子标记物分型、指导治疗和预测预后提供了一种全新的无创影像学评价新方法。

胰腺癌（pancreatic carcinoma，PC）发生、发展、转移和预后与某些特定的分子密切相关。但是，要获得这些分子标记物信息，需要取得胰腺组织或胰液。术前获取肿瘤组织需要通过穿刺活检或胰液细胞学检查，它们具有侵入性、获取组织有限、并发症多、受操作者和肿瘤部位影响大，运用于临床受到极大限制。传统影像学方法对PC的诊断主要在解剖学和形态学层面，很难与分子表达建立联系。目前急需一种既无创便捷又准确高效的诊断方法在术前获取PC的分子标记物信息。. 基于此本项目进行了下列研究：首先我们提取胰腺癌患者术前影像组学特征，进一步对特征进行降维和选择，发现部分组学特征和PC淋巴结转移高度相关，建立的回归预测模型能够准确的预测术后淋巴结转移信息，模型在训练集和测试集的曲线下面积分别为0.75和0.81。其次，我们发现部分影像组学特征与术后胰腺癌切缘情况高度相关，建立的动脉期组学分值对胰腺癌SMV切缘的预测诊断的曲线下面积为0.838，敏感度为77.8%，特异度为75.6%，准确率为76.24%。最后，我们对胰腺癌间质中最重要的成分成纤维细胞活化蛋白（FAP）进行预测。我们通过量化整张病理组织切片中FAP的表达，并和生存建立联系，将患者分为FAP低表达组和FAP高表达组。对患者术前CT 和平扫MRI各个扫描序列图像高通量提取影像组学特征。利用提取的影像组学特征建立机器学习的预测模型。基于CT图像建立的XGBoost分类器和基于MRI建立的多层感知神经网络分类器对FAP的预测均表现了良好的预测能力。建立的XGBoost分类器训练集AUC为 0.97，验证集AUC为 0.75。敏感度、特异度、准确率、阳性预测值和阴性预测值在训练集为97.22%，96.55%，0.97，0.95和0.98，在测试集为64.00 %，90.91%，0.79，0.84和 0.77。建立的MLP神经网络分类器训练集AUC为 0.84，验证集AUC为 0.77。敏感度、特异度、准确率、阳性预测值和阴性预测值在训练集为75.00%，79.41%，0.77，0.86 和0.66，在测试集为85.00 %，63.16%，0.74，0.71和 0.80。本项目完成为胰腺癌诊治精准医学的发展提供无创评价的新方法。