深圳大学学报理工版

髋关节发育性不良(developmental dysplasia of the hip, DDH)是常见的先天性关节疾病之一.目前临床上常采用Graf法对婴儿进行髋关节超声筛查,以提早发现病情,提高治愈率.Graf方法高度依赖标准切面的选取和关键解剖结构的识别,对医生的知识和经验要求较高.提出智能化婴儿DDH辅助筛查系统,建立自动化筛查流程,实现自动识别标准切面并测量发育指标α角和β角.标准切面自动识别模块基于少样本单类别分类(few-shot one-class classifier, FOC)的神经网络,通过自监督训练方式学习标准切面的特征信息,预测图像的标准化得分. α角和β角快速测量模块基于快速实例网络(fast instance network, FIN),通过高效的单阶段的网络架构和多任务学习模式,对标准切面实时测量并将结果可视化.分别通过自动识别标准切面和自动测量发育指标两类实验对所建立的辅助筛查系统进行验证.与单类别支持向量机、深度支持向量数据描述网络和Ganomaly网络相比,FOC方法的接受者操作特征曲线下面积达到76.43%,性能最优; 与全卷积网络、Unet和deeplab V3相比,采用FIN模块测量的α角和β角的平均绝对误差分别为2.48°和4.38°,推理速度达到33.88帧/s,速度最快且性能最优.实验结果表明,该系统可降低对训练数据量的依赖,有助于提升DDH临床筛查的同质化水平,控制测量质量,提高临床筛查的工作效率.

Developmental dysplasia of the hip(DDH)is one of the common congenital joint diseases. In order to find the disease as early as possible and thus improve the cure rate, the Graf method is often used in the clinical diagnose of DDH by early ultrasound screening of hip joints of infants. The method highly depends on the selection of standard planes and identification of key anatomical structures, which requires the high level knowledge and experience of physicians. This study proposes an intelligent infant DDH auxiliary screening system. It is designed to realize the automatic identification of standard planes and measure two indicators, i.e., α angle and β angle. A few-shot one-class classifier(FOC)is used as the automatic recognition standard plane module which can recognize the standard planes from hip joint ultrasound videos, learn the features of standard planes with self-supervised training, and predict the image standardized score. A fast instance network(FIN)is served as the automatic fast measurement module which measures two angles on the standard plane in real-time and visualizes the results through the one-stage architecture and multi-task learning mode. The performance of our system is verified with two kinds of experiments: automatic identification of the standard plane and automatic measurement of development indicators. Compared with other one-class classifiers, such as one-class support vector machine, deep support vector data description, and Ganomaly, the proposed method obtains better prediction performance with the area under the receiver operating characteristic curve 76.43%. Compared with other common neural networks, such as full convolutional network, Unet and deeplab V3, the FOC method obtains better prediction performance with lower computation complexity, i.e., the average errors in α and β angles are 2.48° and 4.38° and the inference speed is 33.88 frames/s. The experimental results show that the system can reduce the dependence on amount of training data, improve the homogeneity level of clinical screening for DDH, control the quality of measurement, and improve the efficiency of clinical screening.

引言
1 智能化DDH辅助筛查系统
2 实验及结果分析

图1 Graf法临床标准<br/>Fig.1 Clinical standards of Graf method

图1 Graf法临床标准
Fig.1 Clinical standards of Graf method

图2 智能化DDH辅助筛查系统流程图<br/>Fig.2 Flow chart of the intelligent auxiliary screening system of DDH

图2 智能化DDH辅助筛查系统流程图
Fig.2 Flow chart of the intelligent auxiliary screening system of DDH

图3 自动识别标准切面模块的网络框架<br/>Fig.3 The framework of the standard plane automatic recognition network

图3 自动识别标准切面模块的网络框架
Fig.3 The framework of the standard plane automatic recognition network

图4 自动快速实例网络结构<br/>Fig.4 The framework of the fast instance network

图4 自动快速实例网络结构
Fig.4 The framework of the fast instance network

表1 标准切面分类数据和自动测量实验集分布<br/>Table 1 Dataset of stand plane classification and automatic measurement个

表1 标准切面分类数据和自动测量实验集分布
Table 1 Dataset of stand plane classification and automatic measurement个

表2 少样本单分类网络数据集划分<br/>Table 2 Dataset partition of few-shot one class classifier

表2 少样本单分类网络数据集划分
Table 2 Dataset partition of few-shot one class classifier

表4 不同单分类网络性能的定量评估结果1),2)<br/>Table 4 Quantitative evaluation results of different classification networks%

表4 不同单分类网络性能的定量评估结果1),2)
Table 4 Quantitative evaluation results of different classification networks%

图5 不同分类网络ROC曲线对比<br/>Fig.5 Comparison of receiver operating characteristic curve of different classification networks

图5 不同分类网络ROC曲线对比
Fig.5 Comparison of receiver operating characteristic curve of different classification networks

图6 FOC网络分类结果示例<br/>Fig.6 Examples of classification results of FOC network

图6 FOC网络分类结果示例
Fig.6 Examples of classification results of FOC network

表5 不同网络目标检测及分割性能的定量评估结果1)2)<br/>Table 5 Quantitative evaluation results of object detection and segmentation performance of different networks

表5 不同网络目标检测及分割性能的定量评估结果1)2)
Table 5 Quantitative evaluation results of object detection and segmentation performance of different networks

图7 不同网络分割结果定性展示<br/>Fig.7 Qualitative display of different network segmentation results

图7 不同网络分割结果定性展示
Fig.7 Qualitative display of different network segmentation results

表6 不同网络自动测量结果与标准值的MAD1)<br/>Table 6 MAD between automatic measurement results and standard values of different networks

表6 不同网络自动测量结果与标准值的MAD1)
Table 6 MAD between automatic measurement results and standard values of different networks

图8 FIN网络自动测量角度结果可视化示例<br/>Fig.8 Example of FIN network automatic angle measurement results

图8 FIN网络自动测量角度结果可视化示例
Fig.8 Example of FIN network automatic angle measurement results

[1] KANG Y R,KOO J. Ultrasonography of the pediatric hip and spine[J]. Ultrasonography, 2017, 36(3): 239-251.
[2] SHORTER D, HONG T, OSBORN D A. Cochrane review: screening programs for developmental dysplasia of the hip in newborn infants[J]. Evidence-Based Child Health: a Cochrane Review Journal, 2013, 8(1): 11-54.
[3] KOCER H E, CEVIK K K, SIVRI M, et al. Measuring the effect of filters on segmentation of developmental dysplasia of the hip[J]. Iranian Journal of Radiology, 2016, 13(3):e25491.
[4] GRAF R. Hip sonography: diagnosis and management of infant hip dysplasia[M]. Berlin: Springer Science & Business Media, 2006.
[5] QUADER N, HODGSON A J, MULPURI K, et al. Automatic evaluation of scan adequacy and dysplasia metrics in 2-D ultrasound images of the neonatal hip[J]. Ultrasound in Medicine & Biology, 2017, 43(6): 1252-1262.
[6] CHEN Hao, WU Lingyun, DOU Qi, et al. Ultrasound standard plane detection using a composite neural network framework[J]. IEEE Transactions on Cybernetics, 2017, 47(6): 1576-1586.
[7] CHEN Hao, DOU Qi, NI Dong, et al. Automatic fetal ultrasound standard plane detection using knowledge transferred recurrent neural networks[C]// Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention. Berlin: Springer, 2015,9349: 507-514.
[8] YU Zhen, NI Dong, CHEN Siping, et al. Fetal facial standard plane recognition via very deep convolutional networks[C]// The 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society.[S. l.]: IEEE, 2016: 627- 630.
[9] QUADER N, HODGSON A, MULPURI K, et al. Automatic assessment of developmental dysplasia of the hip[C]// The 12th International Symposium on Biomedical Imaging. New York, USA: IEEE, 2015: 13-16.
[10] LONG J,SHELHAMER E,DARRELL T.Fully convolutional networks for semantic segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA: IEEE, 2015: 3431-3440.
[11] RONNEBERGER O, FISCHER P, BROX T. U-net: convolutional networks for biomedical image segmentation[C]// International Conference on Medical Image Computing and Computer-assisted Intervention. Berlin: Springer, 2015: 234-241.
[12] EL-HARIRI H, MULPURI K, HODGSON A, et al. Comparative evaluation of hand-engineered and deep-learned features for neonatal hip bone segmentation in ultra-sound[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Berlin: Springer, 2019:12-20.
[13] ZHANG Zichen, TANG Min, COBZAS D, et al. End-to-end detection-segmentation network with ROI convolution[C]// The 15th International Symposium on Biomedical Imaging.[S.l.]: IEEE, 2018: 1509-1512.
[14] HAREENDRANATHAN A R, ZONOOBI D, MABEE M, et al. Toward automatic diagnosis of hip dysplasia from 2D ultrasound[C]// The 14th International Symposium on Biomedical Imaging. Melbourne, Australia: IEEE, 2017: 982-985.
[15] SEZER A, SEZER H B. Deep convolutional neural network-based automatic classification of neonatal hip ultrasound images: a novel data augmentation approach with speckle noise reduction[J]. Ultrasound in Medicine and Biology, 2020, 46(3): 735-749.
[16] KHAN S S, MADDEN M G. A survey of recent trends in one class classification[C]// Irish Conference on Artificial Intelligence and Cognitive Science. Berlin: Springer, 2009: 188-197.
[17] CHEN Yunqiang, ZHOU Xiang, HUANG T S. One-class SVM for learning in image retrieval[C]// Proceedings of International Conference on Image Processing.[S. l.]: IEEE, 2001, 1: 34-37.
[18] RUFF L, VANDERMEULEN R A, GÖRNITIZ N, et al. Deep one-class classification[C]// International Conference on Machine Learning.[S. l.: s. n.], 2018: 4393- 4402.
[19] GOLAN I, EL-YANIV R. Deep anomaly detection using geometric transformations[EB/OL].(2018- 05-28)[2021- 02- 04]. https://arxiv.org/abs/1805.10917.
[20] HENDRYCKS D, MAZEIKA M, KADAVATH S, et al. Using self-supervised learning can improve model robustness and uncertainty[EB/OL].(2019- 06-28)[2021- 02- 04]. https://arxiv.org/abs/1906.12340.
[21] KINGMA D P, WELLING M. Auto-encoding variational bayes[J].(2013-12-20)[2021- 02- 04]. https://arxiv.org/abs/1312.6114.
[22] AKCAY S, ATAPOUR-ABARGHOUEI A, BRECKON T P. GANomaly: semi-supervised anomaly detection via adversarial training[C]// Asian Conference on Computer Vision. Berlin: Springer, 2018: 622- 637.
[23] ZONG Bo, SONG Qi, MIN M R, et al. Deep autoencoding Gaussian mixture model for unsupervised anomaly detection[C]// International Conference on Learning Representations. Vancouver, Canada:[s. n.], 2018: 1-19.
[24] HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway, USA: IEEE, 2016:770-778.
[25] HE Kaiming, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice, Italy: IEEE, 2017: 2961-2969.
[26] BOLYA D, ZHOU Chong, XIAO Fanyi, et al. Yolact: real-time instance segmentation[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seouul: IEEE, 2019: 9157-9166.
[27] LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA: IEEE, 2017: 2117-2125.
[28] LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]// European Conference on Computer Vision. Berlin: Springer, 2016: 21-37.
[29] QUADER N, HODGSON A J, MULPURI K, et al. Automatic evaluation of scan adequacy and dysplasia metrics in 2-D ultrasound images of the neonatal hip[J]. Ultrasound in Medicine & Biology, 2017, 43(6): 1252-1262.

备注

引言

1 智能化DDH辅助筛查系统

2 实验及结果分析

补充材料文件请点击下载

期刊信息