基于多尺度卷积特征融合的肺结节图像检索方法

为了解决肺结节图像检索中特征提取难度大、检索精度低下的问题，提出了一种深度网络模型——LMSCRnet用于提取图像特征。首先采用多种不同尺寸滤波器卷积的特征融合方法以解决肺结节大小不一引起的局部特征难以获取的问题，然后引入SE-ResNeXt块来得到更高级的语义特征同时减少网络退化，最后得到肺结节图像的高级语义特征表示。为满足现实中大数据量检索任务的需求，将距离计算及排序过程部署到Spark分布式平台上。实验结果表明，基于LMSCRnet的特征提取方法能够更好地提取图像高级语义信息，在肺结节预处理数据集LIDC上能够达到84.48%的准确率，检索精度高于其他检索方法，而且使用Spark分布式平台完成相似度匹配及排序过程使得检索方法能够满足大数据量检索任务需求。

Retrieval method of pulmonary nodule images based on multi-scale convolution feature fusion

In order to solve the difficulty of feature extraction and low accuracy of retrieval in pulmonary nodule image retrieval， a deep network model named LMSCRnet was proposed to extract image features. Firstly， the feature fusion method of convolution of filters with different scales was adopted to solve the problem of difficulty in obtaining local features caused by different sizes of pulmonary nodules. Then， the SE-ReSNeXt block was introduced to obtain the semantic features with higher level and reduce network degradation. Finally， the high-level semantic feature representation of pulmonary nodule image was obtained. In order to meet the needs of massive data retrieval tasks in real life， the distance calculation and sorting process were deployed on the Spark distributed platform. The experimental results show that the feature extraction method based on LMSCRnet can better extract the image high-level semantic information， and can achieve 84.48% accuracy on the preprocessed dataset of lung nodules named LIDC， and has the retrieval precision higher than other retrieval methods. At the same time， using Spark distributed platform to complete similarity matching and sorting process enables the retrieval method to meet the requirements of massive data retrieval tasks.