MRI Knee Joint Segmentation Research Based On U-Net

As the most human activity and the most loading joint,the integrity of the knee joint is an important part of people’s happy life.The diagnosis of the knee joint mainly relies on doctors to analyze MRI images.The use of artificial intelligence technology to perform pixel-level segmentation of medical images,which plays an important role in improving diagnostic efficiency.In the current MRI knee image segmentation,the U-Net network using the coderdecoder This symmetrical structure has achieved good results.However,due to the characteristics of knee tissue,the proportion of different bone tissues in medical images is unbalanced,the gray characteristics between bones and cartilage are similar,the bone tissue is small from the surrounding muscle tissue,tissue fluid,etc.Adhesion and so on.These characteristics cause various problems during the network,and the division effect is not ideal.By proposing a modified U-NET network,this thesis improves the segmentation accuracy of the network of knee tissue.The specific work of this thesis is as follows:First of all,in order to solve the loss of information that accounted for smaller bone and cymbal cartilage due to the reduction of the feature map due to the reduction of the feature map,and then the network was inadequate to learn from knee tissue and cause misalignment between the organization.Based on the U-Net network,this thesis uses a modified Inception module on each layer of the encoder,adding the width of the network,and a multi-scale semantic information,increasing the network’s adaptability to different scale.Under the overall framework of the encoder,a dense connection module idea is used,and the characteristics you learn from each layer of the network during the encoder phase network are used.Secondly,in order to solve the problem of incompleteness and rough edges or even misplaced in the segmentation of networks and tibial.This thesis uses different expansion voltage convolution at each level of the decoder.Under the premise of increasing the calculation amount,it effectively enhances the network experience.At the same time,the use of FPN module ideas on the decoder as a whole,integrates the characteristics from different levels,and increases the segmentation accuracy of the network to occupy a relatively large bone tissue.Again,in order to solve the problem of misunderstanding the muscle tissue as a bone tissue,the network has better guided the learning direction of the network.This thesis uses and uses the generated two-value label and attention mechanism to conduct supervision and learning,which strengthens the attention of the network to the image prospects and the organizations of each part in the prospects,especially to increase the attention of bone tissues that account for smaller bone tissues.At the same time,it also gives different weights to the characteristics of stitching on the channel and achieved good results.Finally,in order to verify the validity of the network modification of this thesis,the ablation experiments were performed first,and the module modified in this thesis can be verified that the segmentation accuracy of the network can be improved to a certain extent.After that,a comparative experiment was conducted,and the U-Net network and the commonly used mainstream U-Net network were selected for comparison.Compared with other networks,the network has an advantage on various indicators,proving the effectiveness of the amendments in this thesis.