基于深度学习算法Mask R-CNN的甲状腺结节检测模型研究  

Study on Thyroid Nodule Detection Model Based on Deep Learning Algorithm Mask R-CNN

作　　者：王杰[1] 王至诚 娄帅 董建成 曹新志 WANG Jie;WANG Zhicheng;LOU Shuai;DONG Jiancheng;CAO Xinzhi(Department of Information,Jiangsu Province Hospital on Integration of Chinese and Western Medicine,Nanjing 210028,China;Jiangsu Zhongkang Software Co.Ltd.,Nantong 226001,China)

机构地区：[1]江苏省中西医结合医院信息中心,南京210028 [2]江苏中康软件有限责任公司,南通226001

出　　处：《医学信息学杂志》2025年第3期84-89,共6页Journal of Medical Informatics

基　　金：国家自然科学基金资助项目(项目编号:81971708)。

摘　　要：目的/意义采用基于区域卷积神经网络的目标掩码分割算法(mask region-based convolutional neural network, Mask R-CNN)建立目标检测模型,智能识别甲状腺超声图像结节位置,为超声医生决策提供参考。方法/过程收集超声结节图像1 650张,使用labelme工具进行结节位置标注。对Mask R-CNN的主干网络分别采用MobileNetV3、ResNet50、ResNet101和ResNet152进行替换,并引入特征金字塔和感兴趣区域对齐,采用迁移学习训练策略训练模型,比较不同网络下目标检测效果。结果/结论主干网络采用ResNet101训练的模型平均精确度为86.8%,平均召回率为95.3%,平均F1分数为90.6%,优于其他主干网络,能更精确地检测甲状腺结节,具有一定临床应用价值。Purpose/Significance To establish an object detection model through mask region-based convolutional neural network(Mask R-CNN),so as to intelligently identify the nodule location in thyroid ultrasound images and provide references for the decision-making of ultrasound doctors.Method/Process 1650 ultrasound nodule images are collected,and the labelme tool is used to label the nodule locations.The backbone network of Mask R-CNN is replaced by MobileNetV3,ResNet50,ResNet101 and ResNet152 respectively,feature pyramid network and region of interest align are introduced.The model is trained using transfer learning training strategy and the object detection performance is compared under different networks.Result/Conclusion The model trained with ResNet101 for the backbone network has an average accuracy of 86.8%,an average recall rate of 95.3%,and an average F 1 score of 90.6%,which is superior to other backbone networks and can detect thyroid nodules more accurately,and has certain clinical application value.

关 键 词：甲状腺结节 Mask R-CNN 目标检测 神经网络 

分 类 号：R-058[医药卫生]