基于迁移学习和三维卷积神经网络检测肺结节

Detection of pulmonary nodules based on transfer learning and three dimensional convolutional neural network

目的采用迁移学习多尺度三维卷积神经网络(3D CNN)检测肺结节。方法综合多种方法分割肺实质区域,提取三维候选肺结节;搭建多尺度特征结构3D CNN模型,通过基于权值的迁移学习方法微调网络模型结构;提取数据集,训练微调后的三维网络模型,以之检测肺结节。结果本方案对大(>15 mm)、中(≥5 mm且≤15 mm)肺结节的平均准确率达97.28%,对小(<5 mm)结节平均准确率达92.31%,综合性能优于传统方法及深度学习方法。结论基于迁移学习和3D CNN可自动精确检测不同大小肺结节。

Objective To detect pulmonary nodules with a multi-scale three dimensional convolutional neural network(3D CNN)based on transfer learning.Methods Pulmonary parenchyma area were segmented with various methods,and 3D candidate pulmonary nodules were extracted.Then multi-scale 3D CNN model of feature structure was built,and the network model structure was fine-tuned using weighted transfer learning method.Datasets were extracted to train the fine-tuned 3D network model,and then pulmonary nodules were detected using this model.Results The average accuracy of this scheme for detecting medium(≥5 mm and≤15 mm)and large(>15 mm)pulmonary nodules was 97.28%,and for small(<5 mm)nodules was 92.31%.The comprehensive performances of this scheme were better than those of traditional methods and deep learning methods.Conclusion Pulmonary nodules could be detected automatically and accurately based on transfer learning and 3D CNN.