基于BCNN的胎儿颅脑超声横切面识别算法

孕期超声检查是评估胎儿大脑发育、检测生长异常的重要步骤,开展对胎儿早期检查准确高效的诊断研究具有重要的临床价值。文中使用双线性卷积神经网络BCNN进行胎儿颅脑横切面识别,提出了BCNN-R和BCNN-S两种算法。BCNN算法首先对输入的胎儿颅脑超声影像数据进行预处理,去除个人信息等敏感信息;其次,利用两路并行的子网络从影像数据中提取辨识度高、鲁棒性强的横切面特征,并将其融合得到有助于识别的细微特征;最后使用线性连接层进行识别和分类。为了验证算法的有效性,在自建胎儿超声数据集JFU19上进行了对比实验,实验结果表明,所提算法相比常见的深度网络(GoogleNet,DenseNet,SeNet等)在分类性能上有明显的提升,其中BCNN-S算法的总体准确率达到了88.95%,BCNN-R在水平横切面的识别上达到了97.22%的精确度和88.61%的召回率。此外,在公开数据集HC18上进行了实验,BCNN算法的准确率、精确度、召回率分别达到了89.48%,87.66%和87.71%,进一步验证了算法的有效性。

Transverse Section Recognition Algorithm Based on BCNN for Fetal Craniocerebral Ultrasound

Ultrasound examination during pregnancy is an important step to discover whether the fetus is abnormal,so it is of great value to carry out accurate and efficient examination and diagnosis on fetus.In this paper,bilinear convolutional neural network(BCNN)is used to identify the transverse section of the fetal head.On this basis,we propose BCNN-R and BCNN-S.The BCNN model takes the fetal craniocerebral ultrasound image as input,and firstly preprocesses the input data.Secondly,the two parallel sub-networks can extract the cross-sectional features with high identification and strong robustness from the input,after that the model integrates the extracted features,which is helpful to extract the fine features for recognition.Finally,the linear layer outputs the result of classification.In order to verify the effectiveness of the proposed algorithm,this paper makes contrast experiment on self-built fetal ultrasound dataset JTU19.The experimental results show that the proposed algorithms have ob-vious improvement on classification performance compared with the basic network(GoogleNet,DenseNet,SeNet,etc.),the overall accuracy of BCNN-S reaches 88.95%,and the precision and recall of BCNN-R in horizontal cross-sections achieves 97.22%and 88.61%.In addition,we also use the public dataset HC18 to conduct classification experiments.The accuracy,precision and recall of BCNN reach 89.48%,87.66%and 87.71%respectively,which further verifies the effectiveness of the proposed algorithms.