Research On Brain Tissue Segmentation And Head Model Reconstruction Based On Deep Learnin

Transcranial Magnetic Stimulation(TMS)is a non-invasive brain Stimulation technique that has great potential in the diagnosis and treatment of neurological and psychiatric diseases.When applying TMS treatment,the coils are usually placed in a fixed position based on the patient’s etiology,and the differences in the patient’s brain anatomy are not considered enough to achieve optimal stimulation.In order to achieve better treatment effect,it is necessary to have a personalized head model of the patient,and then find the best stimulation plan according to the electric field simulation.Head models are commonly generated via the segmentation of magnetic resonance images(MRI)into different brain tissues.The segmentation accuracy of the brain tissue determines the reconstruction effect of the head model.In order to improve the segmentation accuracy of brain tissue and ensure the effect of head model reconstruction,this paper proposes two network models for brain tissue segmentation based on deep learning theory.The automatic segmentation of brain tissue can be completed with a small amount of data,and a higher segmentation accuracy is obtained.Firstly,this paper proposes a multi-scale attention network(MSAN)based on transfer learning.Due to the particularity of medical images,there are less sample data for training,which makes the network easy to produce over fitting.Transfer learning is used to solve the problem of small sample data.Multiple sub-networks are constructed to extract feature from multi-mode MRI,and complementary information of different modes was fully utilized.The designed multi-scale fusion module can obtain the multiscale information of the input data and fuse the high and low receptive fields of the feature map.The CBAM attentional mechanism is used to focus on the important features of multi-scale fusion.The network was verified on the MRBrain S13 challenge data set.The experimental results showed that the Dice evaluation indexes of white matter,gray matter and cerebrospinal fluid were 89.92%,86.64% and 85.27%,the HD evaluation indexes were 1.65,1.34 and 1.82,and the absolute volume difference evaluation indexes were 6.52,5.46 and 6.52,ranking the second place in the challenge.Secondly,aiming at the problem that the model network is large due to the use of transfer learning in MSAN,a multi-scale residual network(MSRN)is proposed based on MSAN.The method of transfer learning is abandoned to reduce network parameters,and use the MRBrain S18 challenge data with more samples to train the network.The designed multi-scale residual module can obtain and fuse the multi-scale features of brain tissue and prevent the degradation of the network.Experimental results on MRBrain S18 show that MSRN has strong segmentation performance and ranks first in the three-category challenge on MRBrain S18.Finally,the segmented brain tissue data was imported into the medical image processing software for 3D reconstruction of the head model.The geometric model of the head model is established,and the mesh optimization and volume mesh division are performed on the basis of the geometric model to realize the reconstruction of the personalized head model for TMS simulation.