目的：胆囊癌作为一种发病率不高但是预后极差的消化道恶性肿瘤，早期诊断肿瘤以及准确评估肿瘤情况对选择治疗方案、改善患者预后有着重要意义。然而，基于传统影像学检查的胆囊癌早期诊断及肿瘤评估一直非常困难。影像组学作为一个近期兴起的研究方法，在肺癌、乳腺癌、神经胶质瘤等肿瘤疾病的诊断中发挥出极佳的诊断效果，例如预测病理类型、预测肿瘤情况、预测患者预后、预测肿瘤基因型甚至预测药物治疗效果。目前，暂无研究使用影像组学预测胆囊癌的病理类型及浆膜层浸润情况，因此，本研究计划使用增强CT进行影像组学特征分析，构建机器学习模型来判断胆囊癌病理类型及浆膜层累及情况。方法：本研究为回顾性研究，共纳入了80名术后石蜡病理诊断为胆囊癌的患者，收集术前增强CT影像及病理结果。进行图像资料的预处理后，由放射科医生划定ROI。影像组学特征参数的提取由Python语言完成。根据病理类型分组，使用独立对象克 鲁沃斯卡尔-沃利斯检验并两两比较，确定病理类型之间存在显著性差异的参数。之后，以浆膜层受累情况作为分组依据，使用独立对象曼-惠特尼U检验确定与浆膜层相关的参数。整合所有有差异的参数，使用Tensorflow线性结构构建3层神经网络 模型，每层 20 个神经节点。以训练集及验证集的曲线下面积稳定且高于 0.75 作为条件筛选模型。最后使用随机抽取的验证集评估模型效果。结果：本研究发现在不同病理类型之间存在显著性差异（p<0.05）的影像组学参数 共计127个，差异参数主要分布在形状特征组（41个）和纹理特征组（75个）。不同病理类型之间两两比较后发现，影像组学参数的差异主要来自高分化腺癌与低分化腺癌（67个）、高分化腺癌与中低分化腺癌（77个）。使用以上差异参数可以构建区分高分化腺癌、低分化腺癌及其他腺癌的机器学习模型，曲线下面积可以稳定在0.8以上。在浆膜层受累情况之间存在显著性差异（p<0.05）的影像组学参数共75 个，使用这些差异参数可以构建预测浆膜层受累情况的模型，曲线下面积可以稳定在 0.75 以上。结论：在不同胆囊癌病理类型和不同浆膜层受累情况之间，增强 CT 的部分影像组学特征参数存在显著性差异。基于这些差异参数，可以构建区分胆囊癌病理类型及浆膜层受累情况的增强CT影像组学的机器学习模型。

Purpose: Gallbladder cancer is a malignant tumor with a low incidence but a very poor prognosis. Early diagnosis of the tumor and accurate assessment of the tumor are of great significance for improving the prognosis of patients. However, the early diagnosis and tumor evaluation based on traditional imaging examinations have been very difficult. As a recently emerging research method, radiomics has played an excellent diagnostic effect in lung cancer, breast cancer, glioma and other tumor diseases, such as predicting pathological types, tumor conditions, patient prognosis, tumors Genotype even the effect of drug treatment. At present, there is no research using radiomics to predict the pathological type and serosal infiltration of gallbladder cancer. Therefore, this research plans to analyze enhanced CT radiomics feature of gallbladder cancer and construct a machine learning model to determine the pathological type and serous membrane involvement. Methods: This study is a retrospective study. A total of 80 patients diagnosed as gallbladder cancer by paraffin pathology after surgery were enrolled, and the preoperative enhanced CT images and pathological results were collected. After preprocessing the image data, the radiologist delineated the ROI. The radiomics feature parameters is extracted by Python. Grouping by pathological types, Kluvoskar-Wallis test was used to determine the parameters with significant differences. After that, based on the involvement of the serosal layer, the Mann-Whitney U test was used to determine the parameters related to the serous layer. All the different parameters were integrated and the Tensorflow linear structure was used to build a 3-layer neural network model with 20 neural nodes in each layer. It was the standard of the well-trained model that the area under the curve of the training set and the validation set was stable and higher than 0.75. Finally, a randomly selected validation set is used to evaluate the effect of the model. Results: This study found a total of 127 radiomics parameters with significant differences (p<0.05) between different pathological types. The difference parameters were mainly distributed in the shape feature group (41) and the texture feature group (75). After pairwise comparison between different pathological types, it is found that the difference in radiomics parameters mainly comes from well-differentiated and poorly differentiated adenocarcinoma (67), well-differentiated adenocarcinoma and moderately-poorly differentiated adenocarcinoma (77). Using the above difference parameters, a machine learning model can be constructed to distinguish well-differentiated adenocarcinoma, poorly-differentiated adenocarcinoma and other adenocarcinomas, and the area under the curve can be stabilized above 0.8. There are a total of 75 radiomic parameters with significant differences (p<0.05) between the involvement of the serosal layer or not. Using these difference parameters, a model for predicting the involvement of the serosal layer can be constructed, and the area under the curve can be stabilized above 0.75. Conclusion: There are significant differences in radiomics characteristic parameters of enhanced CT among different pathological types of gallbladder cancer and different serosal layer involvement. Based on these difference parameters, a machine learning model of enhanced CT radiomics can be constructed to distinguish the pathological type of gallbladder cancer and the involvement of the serosal layer.