Brain Tumor Magnetic Resonance Image Segmentation Based On Improved U-Net Method

The development of artificial intelligence technology has made the application of computer-aided diagnosis increasingly widespread.With the help of clinical medical big data analysis,the efficiency and accuracy of computer-aided diagnosis have also been recognized by the industry.In recent years,with the rapid development of deep learning in image processing applications,the use of deep learning methods to segment medical images has also become a research hotspot,especially image segmentation based on convolutional neural networks has become an important application field of medical image processing.The purpose of medical image segmentation is to segment the lesion area of the specified image.The accuracy of the segmentation directly affects the image recognition effect and also affects the doctor’s surgical treatment plan.Brain tumor is a disease in which the intracranial tissues become cancerous.Its early symptoms are not obvious,and it has not attracted enough attention.Once malignant transformation occurs,treatment is more difficult."Early detection and early treatment" are effective means to prevent the malignant transformation of brain tumors.Clinical diagnosis mainly relies on magnetic resonance imaging detection technology to identify brain tumors.In order to improve the efficiency of brain tumor diagnosis and recognition,brain tumor segmentation is particularly important.Because brain tumors have different shapes and different locations,there are also differences in brain tumors between different patients.This makes it difficult for doctors to ensure segmentation accuracy when manually segmenting brain tumors,and the segmentation efficiency is low.At the same time,brain tumors Segmentation puts forward higher requirements on the professional level of doctors.Using deep learning methods to segment medical images can not only achieve the purpose of automatically segmenting tumor lesions,but also significantly improve the accuracy of segmentation.For this reason,two brain tumor MRI image segmentation models based on improved U-Net are proposed: 1)Aiming at the problems of blurred brain tumor boundaries,low contrast,and slower model training,an improved U-Net segmentation based on residuals is proposed.model.The residual network is introduced on the basis of U-Net,and the convolutional layer in U-Net is replaced with residual blocks.A jump connection is added between two residual blocks at the same level to extract more feature information and improve the training of the model Convergence speed in the process,the residual block can be transmitted without attenuation during back propagation,thereby further improving the training efficiency and segmentation accuracy of the model.2)Aiming at the problem of more redundant information in brain tumor images,an improved U-Net segmentation model based on residual attention is proposed.On the basis of the improved U-Net model based on residual error,the attention mechanism is introduced to assign different weights to different features in the image,and the computing resources are concentrated on the region to be segmented in the image,so as to realize the focus on the brain tumor in the image and eliminate redundancy.Additional information can improve the segmentation accuracy of the model.In order to verify the segmentation effect of the two models,the public brain tumor is used as a data set,and the combination of cross-entropy loss and Dice coefficient difference function is used as the loss function to avoid the problem of imbalanced categories in brain tumor MRI images.The model is trained by minimizing the loss function,and the trained model is used to segment the brain tumor MRI image in the test set.The result obtained is very close to the segmentation result manually annotated by the doctor,which proves the feasibility and effectiveness of the two models.Comprehensive evaluation indicators obtained during model testing,the two models proposed are better than the original U-Net,and the improved U-Net segmentation model based on residual attention has the highest segmentation accuracy.