肺结节智能检测和三维可视化系统设计与实现

Design and realization of intelligent detection and three-dimensional visualization system for pulmonary nodules

摘要:

为了提高肺部疾病识别效率，减少肺结节漏诊率，设计了一套肺结节智能检测和三维可视化系统。方法：构建了一个基于RESNET的深度多通道三维卷积神经网络，根据LUNA16公开数据集的888例患者图像，选择权重参数为α=0.5，γ=2的Focal loss损失函数进行训练，在CT图像上对可疑的肺结节进行检测，采用光线投射算法对检测出的结节区域进行体绘制三维重建。结果：经实验测试，该网络与单通道网络和特征金字塔网络（Feature Pyramid network, FPN）相比，准确度最高，为84.8%，系统能够在230s内自动检测肺结节并完成三维重建，对于分辨率1mm/pixel的CT图像灵敏度在98%以上，用户可在浏览器上查看结节检测结果和三维重建模型。结论：该系统突破了终端设备和地域限制，能够为肺部疾病提供辅助诊断，提高诊断效率。

Abstract:

In order to improve the recognition efficiency of lung diseases and reduce the rate of missed diagnosis of lung nodules, a set of intelligent detection and three-dimensional visualization system of lung nodules was designed. Methods: A deep multi-channel three-dimensional convolutional neural network based on RESNET was constructed. Based on the 888 patient images of the LUNA16 public data set, a Focal loss loss function with α = 0.5 and γ = 2 was selected for training. The suspicious lung nodules are detected, and the ray projection algorithm is used to perform volume rendering three-dimensional reconstruction of the detected nodules. Results: After experimental tests, the network has the highest accuracy compared with the single-channel network and Feature Pyramid network (FPN), which is 84.8%. The system can automatically detect lung nodules and complete 3D reconstruction within 230s. The sensitivity of CT images with a resolution of 1mm / pixel is above 98%. Users can view the nodule detection results and 3D reconstruction models on the browser. Conclusion: The system breaks through the limitation of terminal equipment and area, and can provide auxiliary diagnosis for lung diseases and improve the diagnosis efficiency.