胡鹏辉,王 娜,王 毅,等.基于全卷积神经网络的肛提肌裂孔智能识别[J].深圳大学学报理工版,2018,35(No.3(221-330)):316-323.[doi:10.3724/SP.J.1249.2018.03316]
HU Penghui,WANG Na,WANG Yi,et al.Automatic recognition of levator hiatus based on fully convolutional neural networks[J].Journal of Shenzhen University Science and Engineering,2018,35(No.3(221-330)):316-323.[doi:10.3724/SP.J.1249.2018.03316]
点击复制
基于全卷积神经网络的肛提肌裂孔智能识别
胡鹏辉1,王 娜1,王 毅1,王慧芳2,汪天富1,倪 东1

1)深圳大学医学部生物医学工程学院,广东省生物医学信息检测与超声成像重点实验室,医学超声关键技术国家地方 联合工程实验室,广东深圳518060; 2)深圳大学第一附属医院,深圳市第二人民医院超声科,广东深圳518035

生物医学工程; 女性盆底功能障碍性疾病; 肛提肌裂孔; 图像分割; 卷积神经网络; 自动上下文模型; 条件随机场

Automatic recognition of levator hiatus based on fully convolutional neural networks
HU Penghui1, WANG Na1, WANG Yi1, WANG Huifang2, WANG Tianfu1, and NI Dong1

1)School of Biomedical Engineering, Health Science Center, Shenzhen University, Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging, National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, Shenzhen 518060, Guangdong Province, P.R.China; 2)Department of Ultrasound, Shenzhen No.2 People's Hospital, The First Affiliated Hospital of Shenzhen University, Shenzhen 518035, Guangdong Province, P.R.China

biomedical engineering; female pelvic floor dysfunction(FPFD); levator hiatus; segmentation; deep convolutional neural networks; auto-context; conditional random fields

DOI: 10.3724/SP.J.1249.2018.03316

备注

提出一种智能识别肛提肌裂孔的方法,利用端到端的编码器-解码器结构全卷积神经网络,结合自动上下文模型思想,分割出人体盆底超声图像中肛提肌裂孔,采用全连接条件随机场加强边缘约束,对分割结果实现精细化处理,实现肛提肌裂孔的智能识别.通过对372张盆底超声图像进行智能识别,并与医生手动标注结果对比,两者重合率达到95.16%,优于传统卷积神经网络模型,证实基于上下文及条件随机场的神经网络方法能有效识别肛提肌裂孔,具有重要临床应用价值.

We propose a novel deep learning based method to recognize the levator hiatus. The levator hiatus is segmented from the pelvic floor ultrasonic image by using the encoder-decoder architecture of full convolutional neural networks and combining with the auto-context model in an end-to-end manner. The full connection of conditional random fields is used to strengthen the edge constraint in order to refine the segmentation results.The experimental results on 372 pelvic floor ultrasonic images show that the dice ratio of our automatic recognition, compared with the labels of doctors, reaches about 95.16%. The results demonstrate that the proposed method provides more accurate segmentation than the state-of-the-art methods and have a potential clinical value.

·