基于Faster R-CNN及增强CT图像自动检测肝血管瘤

目的 基于快速区域卷积神经网络（Faster R-CNN）构建肝血管瘤自动检测系统，观察其检出增强CT图像中的肝血管瘤的效能。方法 收集经腹部增强CT诊断的128例肝血管瘤患者、共2 304幅增强CT图像，按8∶2比例将其分为训练集（n=102）和测试集（n=26），分别含1 836幅及468幅增强CT图像。利用Faster R-CNN、针对增强CT图像构建自动检测肝血管瘤系统，基于迁移学习方案，采用Resnet50预训练分类网络作为提取特征模块的基础骨架，以区域提议网络提取训练集增强CT图像特征，以边界框分类回归模块输出预测边框的精确位置坐标和类别的概率分数。训练过程中绘制Loss曲线，评估模型对训练集的训练效果及其稳定性；采用随机梯度下降法作为优化器对参数进行调整，以提升模型性能。通过平均精度均值（mAP）评估系统检出测试集增强CT图像中的肝血管瘤的效能。结果 训练集训练过程损失函数Loss曲线中，自动检测系统呈快速下降趋势，提示模型学习能力良好，预测性能稳定。mAP曲线显示，迭代次数epoch为40~80时，系统对测试集468幅增强CT图像检出肝血管瘤的mAP为0.962~0.973，波动小，提示模型已收敛，自动检测效果良好。结论 基于Faster R-CNN的增强CT图像自动检测系统可有效检出肝血管瘤。

Automatic detecting hepatic hemangioma on enhanced CT images using Faster R-CNN

Objective To construct an automatic detection system of hepatic hemangioma on enhanced CT images using faster regional convolutional neural network (Faster R-CNN), and to observe its performance for detecting hepatic hemangioma. Methods A total of 128 patients (2 304 enhanced CT images) with hepatic hemangioma diagnosed with abdominal enhanced CT were enrolled. The patients were divided into training set (102 cases, 1 836 enhanced CT images) and test set (26 cases, 468 enhanced CT images) at the ratio of 8:2. Faster R-CNN was used to construct an automatic hepatic hemangioma detection system for enhanced CT images. Based on the migration learning scheme, Resnet50 pre-trained classification network was taken as the basic skeleton of feature extraction module, while region proposal net was applied to extract the features of enhanced CT images in training set. Then boundary box classification regression module was performed to output the exact position coordinates and category probability scores of the predicted border. During the training process of training set, the Loss curve was drawn to evaluate the training efficacy and stability of the model, while stochastic gradient descent method was used as the optimizer to optimize parameters so as to improve the performance of the model. The mean average precision (mAP) was used to evaluate the performance of the automatic detection system for detecting hepatic hemangiomas on enhanced CT images in test set. Results Loss curve of the automatic detection system showed a rapid decreasing trend, suggesting that the model had good learning ability and stable prediction performance. When the iteration epoch was 40-80, the mAP of the automatic detection system for detecting hepatic hemangioma on 468 enhanced CT images in test set was 0.962-0.973, with small fluctuation, indicating that the model had converged and its automatic detecting effect was good. Conclusion The constructed automatic detection system could effectively detect hepatic hemangioma on enhanced CT images.