摘要
目的:通过全局注意力多任务网络提升CT图像细小骨折检测的感知,通过多任务实现实例级别细小骨折目标的检测,快速、准确地从大量CT图像中识别并定位骨折,以辅助临床及时开展治疗。方法:引入分组非局部(non-local)网络方法,计算CT图像连续切片任何位置和通道之间的远程依赖关系,将多目标检测模型3D RetinaNet单级检测器与医学图像语义分割(3D U-Net)架构相融合,实现端到端的多任务3D卷积网络,以多任务联合的方式实现对细小骨折的实例级别检测。选择医学图像计算与计算机辅助干预(MICCAI)2020挑战赛提供的肋骨骨折公开数据集(Rib Frac Dataset)600例CT扫描图像,通过5∶1的比例划分为训练集(500例)和验证集(100例),测试多任务3D卷积网络的精度性能。结果:多任务3D卷积网络方法的检测精度性能优于单任务网络FracNet、3D RetinaNet及3D Retina U-Net,其平均精度与3D RetinaNet和3D Retina U-Net网络相比分别高出7.8%和11.4%,且优于3D Faster R-CNN、3D Mask R-CNN两种单任务网络检测方法,平均精度分别高出约6.7%和3.1%。结论:全局注意力多任务网络融合不同模块,对于细小骨折检测性能均有提升,引入分组非局部(Non-local)网络方法能够进一步提升对细小骨折目标的检测精度性能。
Objective:To improve the perception of computed tomography(CT)images in detecting fine fracture through multi-task network of global attention,and to realize the detection of the target of fine fracture at case level through multi-task,and to quickly and accurately identify and locate fracture from a large number of CT images,so as to assist doctors to timely conduct treatment.Methods:A grouped Non-local network method was introduced to calculate the remote dependency relationship between each position of CT image continuous sections and channel.A single-stage detector of multi-objective detection model three dimension(3D)RetinaNet was integrated with the medical image semantic segmentation architecture(3D U-Net).A end-to-end multi-task 3D convolutional network was realized,which realized the detection of case level for fine fracture through multi-task collaboration.Select 600 CT scan images from the Rib Frac Dataset of rib fractures provided by the MICCAI 2020 Challenge,and they were divided into training set(500 cases)and test set(100 cases)as the ratio of 5:1 to test the precise performance of multi-task 3D convolutional network.Results:The precise performance of multi-task 3D convolutional network method was better than that of single-task FracNet,3D RetinaNet and 3D Retina U-Net in detection,which average precision was respectively higher 7.8%and 11.4%than 3D RetinaNet and 3D Retina U-Net.It was better than two kinds of single-task network detection method included 3D Faster R-CNN and 3D Mask R-CNN,and the average precision of that was respectively higher 6.7%and 3.1%than them.Conclusion:The integrated different modules of global attention multi-task network can improve the detection performance of fine fracture.The introduction of grouped Non-local network method can further improve the precise performance for the targets of fine fractures in detection.
作者
李瑞瑞
杨晓光
孙世豪
季尚蔚
Li Ruirui;Yang Xiaoguang;Sun Shihao;Ji Shangwei(Beijing FuTong Technology Co.Ltd,Beijing 100020,China;Retirement Office,Beijing Tiantan Hospital,Capital Medical University,Beijing 100070,China;Department of Trauma Orthopedics,Beijing Jishuitan Hospital,Beijing 100035,China)
出处
《中国医学装备》
2024年第3期12-18,共7页
China Medical Equipment
基金
国家重点研发计划“基础科研条件与重大科学仪器设备研发”专项(2021YFF0704100)
