Research On Spinal MRI Image Segmentation Algorithm Based On Convolutional Neural Network

Nowadays,in clinical practice,we often encounter a variety of spinal variation,such as transitional vertebra and vertebral block,which poses a serious threat to people’s health.It is necessary to thoroughly understand these variations in order to avoid wrong spinal surgery.It is necessary to accurately calculate the number of vertebral bodies.With the improvement of medical level and the progress of medical imaging technology,the whole spine should be included in the imaging research,and the routine clinical diagnosis of spine pathology should be made.Medical image segmentation plays an important role in the field of clinical medicine.Traditional medical image segmentation methods not only consume time and energy,but also lead to a large number of errors,which are mainly performed by doctors through manual segmentation based on existing experience and knowledge.This traditional segmentation method has certain limitations.Nuclear magnetic resonance imaging(MRI)is very sensitive to check how much moisture content in the organization and change,can clearly show the change of physiological and biochemical metabolism process information,can provide better diagnostic basis for early lesions,and often more effective than CT imaging can,sooner found lesions and don’t like CT produced with damage of ionizing radiation to human body,based on these advantages,is widely used in the imaging of the spine.MRI analysts(radiologists and orthopedic surgeons)can quickly read the lesions from the presented images.Therefore,it is very important to select a good segmentation method in order to more accurately segment the spine from the medical images for the auxiliary doctors of the human body.In this paper,the deep learning method is applied to the segmentation of the spine to overcome the shortcomings of the traditional methods and improve the timeliness and accuracy.The traditional medical image segmentation algorithm uses the gray information of the medical image to confirm and segment the contour.The segmentation accuracy of this algorithm is mainly affected by the selection of the target,the background label and the construction of the energy function.Therefore,it is not suitable for the segmentation of spinal MRI images.Deep learning algorithms play an important role in the medical field and have applied research significance.They perform well in medical image segmentation,recognition and classification.According to the characteristics of spinal magnetic resonance imaging(MRI),this paper proposes an improved neural network to accurately segment the spine through the sharp contrast between the grayscale of the disc and the vertebra in the MRI image.The data set we used was a total of 210 adult spinal magnetic resonance(MRI)images from the National Medical Innovation Design Competition in 2019.Among the 210 adult spinal MRI images,195 were used as the training set and 15 were used as the test set.The following studies on spinal magnetic resonance images(MRI)were done mainly based on U-NET network:We adopt the gaussian filtering method for spinal nuclear magnetic resonance image(MRI)filtering noise reduction processing,in order to expand the experimental data set,we do to image rotation,cutting,data such as image enhancement method,in order to guarantee the experimental training speed quickly and smoothly,we do to data normalized and standardized operation.We according to the characteristics of the data sets used in building the U-net,residual network(Resnet),Unet++ network,improved Res Unet++ network of spine,nuclear magnetic resonance image(MRI)segmentation,comparing the different network to the spine of nuclear magnetic resonance image(MRI)segmentation accuracy,analysis of the experimental results,determine the most suitable network of nuclear magnetic resonance image(MRI)segmentation of the spine.Unet++ network through sampling,sampling and jumping on a connection,deepen the U-net of network structure,the polymerization characteristics of the scale of the different decoding subnetwork,forming flexible characteristics of the polymerization method,we improved Res Unet++ network using the residual block,atrus space pyramid pool(ASPP),compression and incentives,pay attention to the block,its residual block transmission of information in the layer,allowed to build up a deeper level of neural network to solve the problem of each encoder degradation,improve the interdependencies between the information channel,also cut the cost of computing at the same time,adopt mixing loss function,The higher learning rate is used to train the network,to speed up the learning speed of the network,so as to accelerate the convergence speed of the network,and to improve the segmentation accuracy of the network.In this paper,the improved Resunet ++ network segmentation results show that this segmentation method produces up to 89% segmentation accuracy,strong robustness,segmentation accuracy is relatively stable,and further improve the segmentation accuracy.