Research On The Segmentation Algorithm Of Fetal Cerebellar Ultrasound Images Based On Convolutional Neural Network

With the adjustment of the national fertility policy,the number of elderly pregnant women has increased significantly,and the risk of birth defects has also increased.It is of great significance to achieve accurate monitoring of fetal development and improve the detection rate of fetal congenital malformations for eugenics and improving the quality of the population.Ultrasound detection is an essential part of prenatal examination,but ultrasound images,accompanied by noise,artifacts,and other factors,pose challenges to doctors’ diagnosis.Therefore,how to use deep learning technology to achieve accurate segmentation of fetal organs and tissues under the background of noise interference is the key problem to be solved urgently,and is also the premise of early screening,early diagnosis and early treatment.The existing image segmentation algorithms based on deep learning have few applications in the field of fetal cerebellar image segmentation,and the accuracy is not high,which cannot provide doctors with accurate and reliable basis for diagnosis.In view of this,this thesis will carry out the research of segmentation algorithm of fetal cerebellar ultrasound images based on convolution neural network.Specifically,the research contents and innovative achievements of this thesis are as follows:(1)FCRB U-Net(Fully Connected Residual Block U-Net)cerebellar segmentation algorithm for fetal ultrasound images.To mitigate the loss of feature information in convolution neural network,a network algorithm called FCRB U-Net,which can retain as much semantic information as possible,is constructed for the segmentation of fetal cerebellar ultrasound images.This algorithm starts with the network module and replaces the double convolution module in the original network with the fully connected residual block,which can make up for the loss of feature information in the convolution process.In addition,FCRB U-Net adds the feature reuse module at the decoder to realize the propagation and reuse of deep semantic information.At the same time,in the face of the special background noise and tissue artifact interference in the ultrasound images,FCRB U-Net embeds the efficient channel attention(ECA)module to weight the features of different channels and enhance the effective feature representation.Under the supervision of ECA module,the algorithm has achieved higher segmentation accuracy in fetal cerebellar ultrasound image segmentation.(2)DA-Net(Deep Attention Network)segmentation algorithm for biomedical images.To solve the problem of generalization ability of the network,an image segmentation algorithm called DA-Net,which is universal in a wide range of biomedical image segmentation tasks,is built.This algorithm adds a new layer to the original U-Net to form a six-layer U-shaped network,which increases the depth of the network and makes the network obtain a larger acceptance area.From the aspect of attention mechanism,an improved triple attention module is proposed,which makes full use of the the cross-dimensional interactive information between channel dimension and spatial dimension to enhance the feature representation of valuable information.Then the improved triple attention module and the residual concatenate block are packaged into a composite module,which is to replace the dual convolution module in the original U-Net to help DA-Net obtain stronger feature extraction capability.Finally,the network model is trained by pixel position aware loss.The trained DANet model can not only achieve high segmentation accuracy,but also obtain great generalization ability and robustness.