Alternate Optimization Method for 3D Pulmonary Nodules Retrieval Based on Medical Sign

Abstract

In order to solve the problems such as the complicated process of manual diagnosis and retrieval, high misdiagnosis rate, large amount of data, sparse Hash codes, a 3D ResNet network based on multi-label semantic supervision was proposed to quantify the medical signs of pulmonary nodules and construct a multi-label data set. 3D lung nodules were constructed by trilinear interpolation method, and loss functions were designed by similarity measurement for 3D feature learning. Then Hash codes were constructed. An alternate minimization optimization method was proposed to solve the problem that the traditional method cannot be used because of the discrete Hash code, and the closely expressed Hash code was learned. Finally, a multi-level lung cancer image retrieval algorithm was proposed. The experimental results show that the average accuracy is improved by 18.5% by using the 3D feature proposed in this paper, and the average retrieval accuracy reaches 94.83% on the expanded public data set and the cooperative hospital data set.

Keywords

• lung nodules
• 3d features
• multi-label
• image retrieval
• medical signs