Study On Segmentation Of Long And Narrow Area Based On ResU-Net And Its Application In Knee Cartilage

The knee cartilage acts as a bond in the knee joint.The cartilage lesion is a very serious disease.The precise segmentation of the knee cartilage could help doctors diagnose the extent of cartilage damage.However,due to the individual differences of knee cartilage and the different morphology caused by pathological reasons,the traditional segmentation method has limitations in the task of cartilage segmentation.Cartilage is longer and thinner than normal cartilage due to bone and joint diseases,which may result in segmentation fracture and poor continuous segmentation effect.Moreover,due to the characteristics of the knee joint,small cartilage and large bone are adjacent,the sample is not balanced,and the network is prone to misclassification;In addition,the cartilage edge caused by cartilage injury is not smooth,resulting in poor edge segmentation.To solve the above three problems,this thesis proposes an improved Att Res U-Net model based on Res U-Net to perform the task of knee cartilage segmentation.The specific content of this thesis is as follows:Firstly,to solve the problem of segmentation and fracture of cartilage in segmentation,this thesis proposes a network model based on Res U-Net network to integrate Spatial Attention Layer into the network.The introduction of this module can make the network ignore noise.Paying more attention to more meaningful cartilage areas in space mitigated the continuous segmental rupture to some extent.Secondly,aiming at the problem of misclassification in segmentation,this thesis proposes a deep learning segmentation algorithm that integrates Context Aggregation Attention Module with the help of the attention mechanism.The context attention module uses the location attention module to selectively aggregate the context according to the spatial attention graph and detailed structure at the bottom of the encoder layer,so that the network can obtain richer context correlation and more effective features.It focuses on key pixels,and the combination of the spatial attention module effectively solves the problem of misclassification and misclassification.Thirdly,to solve the problem of poor segmentation effect on the edges of damaged cartilage,this thesis improved the feature fusion strategy of Res U-Net by adding cascade to combine low-level semantic information and high-level semantic information in the image.In this way,the feature map obtained after fusion not only contains richer semantic information,but also includes more semantic information.It also contains the information lost due to continuous upsampling,which effectively solves the problem of poor edge segmentation in the segmentation of damaged cartilage edges by the network.Finally,in order to verify the effectiveness of the proposed algorithm,an ablation experiment was conducted on the MRI data set of human knee joint.Through the ablation experimental data,it can be concluded that each module improved in this thesis can improve the segmentation accuracy of cartilage to a certain extent.In addition,a comparison experiment was conducted with other mainstream algorithms on the human knee MRI data set and SKI10 knee data set.The experimental results show that all indexes of Att Res U-Net proposed in this thesis have obvious advantages over other algorithms,again proving the effectiveness of the proposed algorithm,which is more suitable for the segmentation of medical images in narrow and long areas.