Medical Image Segmentation Based On Convolutional Neural Network

In recent years,because convolutional neural networks have shown good performance in the processing of natural semantic images,researchers have tried to apply them to the field of medical images.Medical images are susceptible to interference from grayscale,targets and noise,and the images will show more prominent characteristics.It is difficult to accurately and effectively segment the images using traditional medical image segmentation algorithms.Based on the characteristics of medical images,this paper discusses the problem of medical image segmentation based on convolutional neural networks.The main work is as follows:1.The study analyzed the semantic segmentation performance of Unet on liver CT images.According to the characteristics of liver CT images,a batch normalization algorithm is added to the network,and the Dice coefficient loss function is used to train the Unet network.The liver data set is divided into training set,validation set and test set for training separately.Comparing the liver segmentation results on Unet with the FCN segmentation results,the IOU value is 0.56 higher than FCN.2.The segmentation of retinal blood vessel images based on improved Unet is studied.In order to solve the problem of imprecise segmentation of blood vessels due to easy changes in a noisy background,an improved Unet model with an attention mechanism is proposed.First of all,in the up-sampling and down-sampling part,the original convolution module is changed to residual convolution,which not only alleviates the problem of gradient explosion or disappearance due to the excessively deep network layers,but also superimposes the features of different levels.Enrich the feature information of the image;secondly,introduce an adaptive attention mechanism in the network structure to make the attention coefficient specific to the local area.By reducing the noise weight,it can solve the problem that a small number of background noises block the blood vessel pixels in the segmentation result The problem,and ultimately achieve better segmentation performance.Compared with Unet,the improved network improves the Dice coefficient on the DRIVE data set by 0.012,and the Dice coefficient on the STRAE data set by 0.01.3.A brain tumor image segmentation algorithm based on improved Unet is studied.Aiming at the segmentation accuracy problems caused by the lack of contextual information connection and the deep network depth in the medical image coefficient segmentation,an improved Unet brain tumor image segmentation algorithm with added dense connections is proposed.This method nests residual connections and jump connections to form a deep supervision network model,changes the jump connections in Unet to multiple types of dense jump connections,and reduces the semantic gap between the encoding path and the decoding path feature map.The attention mechanism and soft thresholding function are added to the improved residual module to effectively reduce image noise and prevent network gradient dispersion.The final experimental results of the algorithm increased the Dice coefficient in the whole tumor area by 0.026,the Dice coefficient in the core area of the tumor increased by 0.018,and the Dice coefficient in the tumor-enhanced area increased by 0.023.