Medical Image Segmentation Based On Deep Learning

In recent years,medical images are more and more widely used in disease diagnosis with its rapid development.Medical image segmentation is one of the vital steps in medical image analysis,and its segmentation results directly affect the subsequent analysis process.Therefore,it is particularly important to segment medical images accurately and efficiently.Glioma is one of the common brain tumors in the brain,and its size,shape and location will vary from different patients.In addition,magnetic resonanc images sometimes have some problems such as uneven grayscale.Therefore,the automatic segmentation of magnetic resonanc images of gliomas has always been a hot and difficult point in this field.The research data set of this thesis is the magnetic resonance image of brain glioma provided by the Department of Nuclear Magnetic Resonance（NMR）of the Lanzhou University Second Hospital.Since there are relatively few high-quality labeled images,this thesis uses the method of data expansion to expand the training set,and a total of 1386 labeled images are obtained.First of all,this thesis tests the previously proposed Unet network model and its changed model,and finds that the network model with the best segmentation result is Attention-Unet,but its segmentation effect can be further optimized.Therefore,this thesis puts forward the Twins-Unet network model on the basis of Unet network model,and studies the influence on the model when data augmentation,adding SE block and Dropout layer and choosing different iterative optimizer.The results show that data augmentation,adding SE block and Dropout layer are meaningful to improve the segmentation effect of the model.The final segmentation model is tested on the test set,and the segmented DSC value,IoU value and loss value are 92.58%,86.20%and 0.0394 respectively.Compared with the Attention-Unet model,the segmented DSC value and IoU value are increased by 3.88%and 5.23%,respectively,and the loss value is reduced by 0.0172.According to the intuitive analysis of the segmentation results of the model,we can see that when the boundary of glioma is obvious and the gray distribution is uniform,the effect of several segmentation models is not much different;but when the boundary of glioma is blurred and the gray distribution is uneven,the segmentation result of Twins-Unet is better than Attention-Unet,so the Twins-Unet segmentation model proposed in this thesis has better segmentation effect.