Research On Placental Segmentation Of Magnetic Resonance Images Based On Deep Learning

The placenta can provide nutrients to the fetus,excrete metabolic wastes from the fetus,complete the material exchange between the fetus and the mother during pregnancy.It has a certain defense against bacteria and pathogens.If the placenta is abnormal during pregnancy,it may lead to the occurrence of related diseases,such as twin-twin transfusion syndrome,invasive placenta and other related diseases,which endanger the health of the mother and the fetus.Therefore,the accurate prenatal detection of placental abnormalities is essential for the early diagnosis and effective treatment of maternal and fetal-related diseases.The quantitative assessment of the placenta based on the acquired prenatal placental images first requires the accurate segmentation of placental tissue.However,the location,size,shape,and thickness of the placenta vary greatly among different pregnant women,making it very challenging to accurately segment placental tissue.The traditional manual segmentation method relies on the professional knowledge and the clinical experience of doctors,which is subjective,This method consumes a lot of manpower and material resources,and the segmentation efficiency is low.The semi-automatic placental segmentation methods still require manual interaction,which easily lead to inconsistent segmentation results and further affect the following analysis.At present,there are few researches and applications of fully automatic placental segmentation methods.Therefore,this paper studied the automatic segmentation method of the placenta based on magnetic resonance image(MRI)to achieve automatic and accurate segmentation of the placenta.At the same time,the lightweight placental fully automatic segmentation model was studied to make it easier to deploy on mobile medical devices,which is beneficial for doctors to diagnose placental related diseases quickly and easily.The specific research contents are as follows:1.In this paper,we proposed a fully automatic placental segmentation method based on deep learning,namely BAA-Net.The network used the BiO-Net as the backbone network.Aiming at the problem that the downsampling operation in BiO-Net will cause the loss of detailed information,which is not conducive to accurately identify the placental boundary.In order to improve the segmentation performance,atrous convolution operations with different dilation rates were used to capture multi-scale features and obtain high-resolution feature maps,which is beneficial to improve segmentation accuracy.In order to make better use of the most useful feature channels,four channel attention modules were introduced in the decoder to highlight the most relevant feature channels,which is beneficial to improve the representation capability of the network.The placenta data of 30 pregnant women,including a total of 1264 2D images,were used for experiments.The obtained placental segmentation Dice value was 0.8385.At the same time,the experimental results show that the method proposed in this paper has good generalization performance when segmenting different placental positions.2.In order to reduce the amount of backbone network parameters and design a more lightweight placental segmentation model,a DSCC-Net network based on the depthwise separable convolution and conditional random field model was designed.In this network,a depthwise separable convolution was used to replace the standard convolution of the backbone to extract image features,and the replaced network was used for placental segmentation to obtain the initial prediction segmentation results.The segmentation accuracy was comparable to the result of original network,but with a smaller amount of parameters and a smaller model size.In order to deal with the discontinuous regions or isolated holes in the initial prediction,the conditional random field model was further used to refine the initial prediction result,and the Dice value was 0.8394.The experimental result indicates that the network has the larger segmentation accuracy and less parameters than the original backbone network.