基于改进Mask R-CNN的多标签甲状腺结节检测模型  

Multi-label thyroid nodule detection model based on improved mask R-CNN

作　　者：吴雯娟 戚琪 邓梓杨 邱桃荣[1] 张卫平 徐盼[2] WU Wenjuan;QI Qi;DENG Ziyang;QIU Taorong;ZHANG Weiping;XU Pan(School of Mathematics and Computer Sciences,Nanchang University,Nanchang 330031,China;Department of Ultrasound,the First Affiliated Hospital of Nanchang University,Nanchang 330006,China)

机构地区：[1]南昌大学数学与计算机学院,江西南昌330031 [2]南昌大学第一附属医院超声科,江西南昌330006

出　　处：《南昌大学学报（理科版）》2023年第2期189-194,共6页Journal of Nanchang University(Natural Science)

基　　金：江西省自然科学基金资助项目(20224BAB216079);江西省重点研发计划(20181BBG70031)。

摘　　要：基于符合中国国情的C-TIRADS甲状腺结节恶性风险分层指南来构建计算机辅助诊断模型具有重要的应用价值,能够提升甲状腺结节和分化型甲状腺癌诊断的规范化、标准化、同质化,提高医生诊断效率和降低劳动强度。本文提出基于深度学习的多标签目标检测模型用以检测识别甲状腺结节、预测甲状腺结节恶性等级以及结节的病理特征,经过在数据集上进行实验对比选取Mask R-CNN作为基准模型,并在基准模型基础上进行优化改进,主要改进包括使用ResNet152-FPN替换原有特征提取网络来提升模型特征能力,设计全新的卷积多标签检测头结构来对结节病理特征进行多标签预测,基于医学先验知识对模型锚框尺寸及比例进行自定义来提升模型的定位精度,最后为模型设计迁移学习训练方案来进一步提升模型性能。实验结果表明,改进模型对甲状腺结节的识别准确率达到了94.4%,对甲状腺结节病理特征的平均识别准确率达到了88.6%。The establishment of computer-aided diagnosis model based on C-TIRADS guidelines for thyroid nodule malignant risk stratification in line with China's national conditions has important application value,which can improve the standardization,standardization and homogeneity of thyroid nodule and differentiated thyroid cancer diagnosis,improve doctors'diagnosis efficiency and reduce labor intensity.In this paper,a multi-label target detection model based on deep learning is proposed to detect and recognize thyroid nodules,predict the malignant grade of thyroid nodules and the pathological characteristics of nodules.Through experimental comparison on the data set,Mask R-CNN is selected as the benchmark model for optimization and improvement,including using ResNet152-FPN to replace the original feature extraction network to improve the model's feature capabilities,design a new convolutional multi tag detector head structure to perform multi tag prediction on pathological features of nodules,customize the size and proportion of the model anchor frame based on medical prior knowledge to improve the positioning accuracy of the model,and finally design a transfer learning training scheme for the model to further improve model performance.The experimental results show that the recognition accuracy of the improved model for thyroid nodules is 94.4%,and the average recognition accuracy for thyroid nodule pathological features is 88.6%.

关 键 词：C-TIRADS 目标检测 甲状腺结节 超声医疗图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] R445.1[自动化与计算机技术—计算机科学与技术]