【摘要】 为辅助临床诊断胆囊癌，使用深度学习技术，通过改进的3D-DenseNet建立一个基于患者增强CT影像的胆囊癌辅助诊断模型。首先，将患者多张动脉期CT转化为三维影像，利用医生标注的胆囊区域，将三维影像切割出感兴趣区域；然后对传统DenseNet网络进行优化，改进Dropout机制与Softmax损失函数并在输出部分将交叉熵函数替换为Focal-loss以进行不平衡校正，从而建立胆囊癌辅助诊断模型；最后，将测试集结果与金标准进行比较，采用ROC曲线、召回率、准确率评估模型性能。通过训练集不断迭代训练，模型损失函数值逐渐收敛，诊断误差也不断下降，在尝试过多种不同的模型结构后，选取出的最优模型准确率为91.4%，特异度为95.2%，灵敏度为88.0%，精确率为95.8%。基于改进3D-DenseNet的胆囊癌诊断模型使用多张患者CT影像数据提取深度特征，充分利用了医疗检测数据，具有较佳的性能和较高的诊断准确率，可以辅助临床进行胆囊癌诊断。更多还原

【Abstract】 To assist in the clinical diagnosis of gallbladder cancer,establish an auxiliary diagnosis model of gallbladder cancer based on patient enhanced CT images through improved 3D-DenseNet using deep learning technology,and discuss its clinical research value. Firstly,multiple arterial CT images of the patient are transformed into three-dimensional images,and the three-dimensional images are cut out of the region of interest by using the gallbladder region marked by the doctor. Then optimize the traditional DenseNet network,improve the Dropout mechanism and Softmax loss function,and replace the cross-entropy function with Focal-loss in the output part to perform imbalance correction,so as to establish the auxiliary diagnosis model of gallbladder cancer. Finally,by comparing the test set results with the gold standard,the ROC curve,recall and accuracy are calculated to evaluate the performance of the model.Through the continuous iterative training of the training set,the value of the model loss function converged and the diagnostic error decreased. After trying a variety of different model structures,the best model was selected,and the accuracy of the optimal model was 91.4%,the specificity was 95.2%,the sensitivity was 88%,and the accuracy was 95.8%.The gallbladder cancer diagnosis model based on 3D-DenseNet uses multiple CT image data of patients to extract depth features,and makes full use of the detection data of patients. The diagnostic model has high performance and accuracy,and can assist the clinical diagnosis of gallbladder cancer.更多还原