Research On Neonatal Brain Magnetic Resonance Image Segmentation Based On Deep Learning

Cerebral palsy refers to a group of movement-limiting disorders with motor and postural development that are caused by non-progressive prenatal,intrapartum or postpartum brain injury.As a non-invasive and non-ionizing radiation detection tool,magnetic resonance imaging(MRI)can provide higher soft tissue resolution and rich image information,which is suitable for the screening of neonatal encephalopathy.However,at present,most primary hospitals in China can only provide conventional MRI mode for neonatal cranial MRI examination.In the face of challenges,the current deep learning-based newborn brain image segmentation algorithms have two shortcomings: 1.over-reliance on data sets;2.the computational complexity of the model increases while the number of network layers is deepened.In order to improve the existing methods of segmentation,this paper is based on the U-Net network,taking T1 and T2 MRI as input to improve the performance of the segmentation model.First,a feature-enhanced dual-modal network is proposed.Secondly,an attention feature enhancement network is proposed based on this network.Finally,through the analysis of comparative experiments,it is concluded that the segmentation effect of the newborn brain MRI segmentation algorithm in this paper is more prominent.The main work of this paper is as follows:(1)A feature-enhanced dual-modal network is proposed for the segmentation of neonatal brain MRI.By designing a dual-channel down-sampling module,the module is composed of dual-channel and maximum fusion,and is used for feature extraction of the input dual-modal image.Maximum fusion is used to fuse the feature images extracted from the subsampling in two channels,which cleverly introduces the complementary advantages between dual-modal image information,and the features extracted through maximum fusion are used to enhance the low resolution of the deep network output Feature images,thereby improving the performance of the segmentation network.Experimental verification shows that the dual-channel down-sampling module can efficiently extract image features.Compared with the typical segmentation network,the feature-enhanced dual-modal network has a better image segmentation effect.(2)To further improve tissue segmentation accuracy in MRI images of newborn brains,it is based on the feature-enhanced dual-modal network.This paper proposes an expansion fusion attention module based on an in-depth understanding of expanded convolution,channel attention mechanism,and spatial attention mechanism.It constructs an attention feature-enhanced network to improve the segmentation performance of the network.Experimental verification shows that the method proposed in this paper can not only use the complementary information between different modalities through fusion and can use the attention module of this paper to learn features with large receptive fields and essential information.A good segmentation effect is obtained.Based on this network model,a segmentation accuracy of 77.28% was achieved on the test set of the d HCP-2017 data set.