Study On Segmentation Model Of Acute Ischemic Stroke Based On Cross-Modality

Acute ischemic stroke is a type of stroke that is a common cerebrovascular disease and accurately identifying and treating ischemic areas in time is the key to reducing the risk of disability and death.At present,magnetic resonance imaging(MRI)is the most commonly used method for clinical diagnosis of stroke.However,the lesions of acute ischemic stroke are characterized by blurred boundaries,more artifacts,and variable size and location on MRI,which leads to cumbersome and time-consuming manual segmentation of the lesions by imaging experts,which leads to manual and segmentation of lesions by imaging experts,resulting in misdiagnosis or missed diagnosis.Some key parameters in traditional image segmentation methods are determined by professional physicians based on their own experience,resulting in rigorous image segmentation results and low accuracy.In recent years,deep learning technology has been widely used to automatically learn image features.This method can effectively assist doctors in improving the accuracy of diagnosis of acute ischemic stroke.This article focuses on the MRI image features of acute ischemic stroke and the image segmentation technology based on deep learning to improve the accuracy of segmentation of acute ischemic stroke lesions.The main research work and contributions are as follows:1.Construct an encoder-decoder segmentation model fused with Cycle GAN.The working principle of the model is as follows: The generated adversarial network Cycle GAN uses the generator and discriminator to realize the cross-mode conversion of CT and MRI images of acute ischemic stroke.The generator makes the discriminator misinterpret the composite image as a real image by means of counter training synthesis,which provides richer semantic information for the encoder and decoder.The encoder uses the spatial spatial pyramid pooling operation to fuse the high-level semantic information of the image,and the decoder extracts the detailed information of the image through upsampling and cascades the high-level information and detailed information of the image with a skip link structure.By comparing thetraditional encoder-decoder model and encoder-decoder segmentation model fused with Cycle GAN,the following conclusions are drawn: Rich input image information can effectively improve the segmentation results of acute ischemic stroke lesions.2.Introduce L2 regularization in Cycle GAN's cyclic uniform loss function to construct a new generation adversarial network I-Cycle GAN and use this as a basis to further improve encoder-decoder segmentation model fused with Cycle GAN,propose an encoder-decoder segmentation model based on I-Cycle GAN.L2 regularization has weight attenuation,on the basis of retaining all the features of the input image,the attenuation coefficient in this direction is determined according to the size of the feature values in each direction of the feature matrix,which reduces the complexity of the network and retains all the information of the input image,and provides richer lesion information for the encoder and decoder.3.The CT and MRI data sets of self-built acute ischemic stroke were compared experimentally to generate confrontation network and segmentation model.The misjudgment rate of I-Cycle GAN composite image was 0.979,which was 0.0732 higher than that of Cycle GAN.In the task of segmentation of acute ischemic stroke,the encoder-decoder segmentation model based on I-Cycle GAN is higher than the traditional encoder-decoder model and encoder-decoder segmentation model fused with Cycle GAN in the three important indicators of Dice coefficient,average cross-combination ratio and sensitivity,respectively 0.886,0.814,0.992.Therefore,the image synthesized by I-Cycle GAN contains richer semantic information,and the encoder-decoder segmentation model based on I-Cycle GAN is more accurate for lesion segmentation.