Research On Key Technologies Of Automatic Segmentation Of Glioma Based On MR Images

Glioma is a common brain tumor disease,and it accounts for 70% of brain tumors.When gliomas develop to a malignant degree,they have a very high mortality rate.Glioma has become one of the major diseases that affect people's health.Magnetic resonance imaging of the brain is a commonly used technique for diagnosing glioma lesions.The delineation of the glioma area plays an important role in the patient's subsequent surgery plan and treatment plan.The delineation of the tumor area mainly depends on the imaging doctor's professional knowledge and manual delineation,but the doctor's manual delineation will inevitably lead to missed diagnosis and misdiagnosis,and manual delineation is very time-consuming and cumbersome.Therefore,computer-assisted automated glioma segmentation is very necessary.The automatic segmentation results of the computer-aided system can not only reduce the workload of doctors,but also provide some assistance for the clinical diagnosis and treatment of gliomas.This paper studies and improves key technologies such as image registration and segmentation modeling in glioma segmentation processing,in order to improve the accuracy and reliability of glioma segmentation.The research work mainly includes the following two aspects:(1)Aiming at the problem that the image to be segmented has spatial deviation and affects the segmentation effect,in the data preprocessing stage,a joint registration algorithm based on 3D multimodal glioma MR image data is proposed.The image registration algorithm can align the deviation between multiple sequences of MR images,so that the data is normalized and the same segmentation result is shared.The registration algorithm proposed in this paper uses Gaussian mixture model to model the pixel intensity and feature points in the image,and then unify the two models into the likelihood function,and finally use the expectation-maximization algorithm to solve the registration parameters.The registration algorithm was first verified in two public datasets RIRE and BRATS.The results show that the registration algorithm proposed in this paper significantly improves the registration accuracy in the case of large deviations,and then applies it to clinical data.The T1 sequence and T2 sequence images are registered,and the offset between the images is significantly reduced.(2)Given the problem that the current glioma segmentation model has insufficient feature learning and no segmentation reliability estimation,a Bayes-Deep Medic-Plus(BDMP model),an automatic glioma segmentation model,is proposed.The network takes image blocks as input,which reduces the complexity of the network and uses a full convolution strategy to achieve pixel-level segmentation.Then,based on Deep Medic,a network model Deep Medic-Plus with improved channel attention mechanism is introduced.Convolutional features of different scales are input to the channel attention module,and all feature maps are connected in series to obtain the final feature map.Based on the Deep Medic-Plus model,this paper introduces the prior distribution to the parameters in the convolution kernel to make the network structure have Bayesian uncertainty.In the solution of the posterior distribution of the weight parameters,this paper combines Monte Carlo sampling Variational inference method.The final segmentation model can evaluate the uncertainty of segmentation results while segmenting gliomas.In the public dataset Bra Ts 2017,the BDMP model proposed in this paper segmented the whole tumor region,and the segmentation result of the whole tumor region is: DICE coefficient of 0.9231,a specificity of 0.9374,and a sensitivity of 0.9021.In clinical data,the segmentation result is a DICE coefficient of 0.7457,specificity of 0.7127,and sensitivity of 0.6938.While segmenting the above two data,this article also analyzes the uncertainty of the segmentation results.In the segmentation results,when the uncertainty value is less than 0.2,the pixel classification accuracy of the two data sets can reach 95.72% and 96.38%.As uncertainty increases,accuracy decreases.This shows that the model proposed in this paper can effectively evaluate the uncertainty of segmentation results.