摘要： 针对实际任务中肺部CT图像标注数据集稀少的问题，提出一种基于自主学习的U-Net模型与C3D多任务学习网络相结合的肺癌诊断方法。对LUNA16数据集和DSB数据集进行预处理，确保切片图像体素、方向一致，利用C3D多任务学习网络模型构建肺结节检测模型，使用165张LUNA16的切片图像和161张DSB的切片图像训练改进的U-Net网络模型，并采用自主学习方式扩充标注样本，构建肿块检测模型。在此基础上，综合结节与肿块检测结果得到最终的肺癌诊断结果。实验结果表明，该方法的肺癌检测精度为85.3%±0.3%，达到了监督学习策略的检测精度。

关键词: 肺结节检测, 肺肿块检测, 肺部CT图像, 自主学习, 多任务学习

Abstract: To address the problem of sparse labeled datasets of lung CT images in actual tasks,this paper proposes a lung cancer diagnosis method that combines the U-Net self-learning model and C3D multi-task learning network.The LUNA16 dataset and DSB dataset are preprocessed to ensure consistent voxels and directions of slice images.Then the method uses the C3D multi-task learning network model to construct a lung nodule detection model,and uses 165 slice images from the LUNA16 dataset and 161 slice images from the DSB dataset to train the improved U-Net network model.The labeled samples are expanded using self-learning to construct a lung mass detection model.On this basis,the node and mass detection results are combined to obtain the final diagnosis of lung cancer.Experimental results show that the lung cancer prediction accuracy of the proposed method is 85.3%±0.3%,which is at the same level of supervised learning method.

Key words: lung nodule detection, lung mass detection, lung CT image, self-learning, multi-task learning

中图分类号: 

• TP391