SPECT骨显像的关节炎病灶自动分割  

Automatic segmentation of arthritis lesions in SPECT bone imaging

作　　者：罗明阳 林强 高瑞婷 满正行 曹永春[2,3] 王海军[4] LUO Mingyang;LIN Qiang;GAO Ruiting;MAN Zhengxing;CAO Yongchun;WANG Haijun(MOE Key Laboratory of China’s Ethnic Languages and Information Technology,Lanzhou 730030,China;School of Mathematics and Computer Science,Northwest Minzu University,Lanzhou 730030,China;Key Laboratory of Streaming Data Computing and Application,Northwest Minzu University,Lanzhou 730124,China;Department of Nuclear Medicine,Gansu Provincial Hospital,Lanzhou 730020,China)

机构地区：[1]西北民族大学中国民族语言与信息技术教育部重点实验室,甘肃兰州730030 [2]西北民族大学数学与计算机科学学院,甘肃兰州730030 [3]西北民族大学动态流数据计算与应用实验室,甘肃兰州730124 [4]甘肃省人民医院核医学科,甘肃兰州730020

出　　处：《现代电子技术》2022年第14期145-152,共8页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61562075);西北民族大学甘肃省一流学科引导专项资金(11080305)。

摘　　要：关节炎是一种常见的多发性生理疾病,临床上易出现骨转移特别是溶骨性转移和关节炎之间的误判。为了从SPECT图像中准确分割关节炎病灶,文中构建一种面向关节炎病灶可靠评估的不同分割模型。首先,对SPECT骨显像数据进行归一化及扩展处理,适度扩充数据量;其次,借助深度学习的特征自动提取功能,构建基于Mask R-CNN的关节炎病灶分割模型,同时构建基于聚类技术的传统分割模型;最后,使用一组真实SPECT骨显像数据,验证所构建的分割模型的有效性。实验结果表明,所构建的深度分割模型可用于关节炎病灶的分割,准确率、召回率和平均交并比分别为0.6044,0.7204和0.6253,聚类分割模型获得的平均交并比为0.802。尽管深度分割模型获得的平均交并比值较低,但其具有实现SPECT图像中关节炎病灶语义分割的潜能。Arthritis is a common and multiple physiological disease,which would easy to be misdiagnosed clinically as bone metastasis especially the osteolytic metastasis.In order to accurately segment arthritic lesions from SPECT images,a different segmentation models for reliable evaluation of arthritic lesions is constructed.The normalization and extension process for the original SPECT imaging data are conducted to moderately expand the data volume.An arthritic lesions segmentation model based on Mask R-CNN is developed by means of the signature automatic generation of deep learning,and meanwhile the traditional segmentation model based on clustering technique is established.A group of real-world bone SPECT imaging data is used to verify the effectiveness of the established segmentation model.The experimental results show that the conducted depth segmentation model can be used to segment the arthritic lesions,and the accuracy rate,recall rate and mean intersection over union(MIoU)are 0.6044,0.7204 and 0.6253,respectively,and the MIoU obtained by the clustering segmentation model is 0.802.Although the MIoU value obtained by the deep segmentation model is low,it has the potential to realize the semantic segmentation of arthritic lesions in SPECT images.

关 键 词：图像分割 功能成像 关节炎 SPECT 数据扩充 Mask R-CNN 深度学习 语义分割 

分 类 号：TN911-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]