Research On Magnetic Resonance Image Segmentation Method Based On Deep Learning And Attention Mechanism

Magnetic Resonance Imaging(MRI)is a widely used imaging technique in clinical medicine,particularly for diagnosing knee joint diseases,due to its significant advantages in soft tissue examination.Currently,high-field and even ultra-high-field MRI techniques are becoming increasingly mature,offering high signal-to-noise ratio and high spatial resolution.However,the excessively high field strength can also lead to local specific absorption rate(SAR)exceeding the limit,causing serious safety issues.Local SAR is difficult to measure and varies greatly among individuals,therefore,it is necessary to build a three-dimensional model of human tissues and then perform electromagnetic simulation to estimate the local SAR.Accurate individual 3D modeling relies on accurate tissue segmentation.Knee joint MRI images are complex and diverse,with some background noise,and limited data,making traditional segmentation methods difficult to use.Currently,deep learning-based methods have been rapidly developing and gaining widespread attention.However,existing deep learning methods still have difficulty achieving satisfactory results for complex knee joint MRI segmentation tasks.In this paper,our research is based on the classic medical image segmentation deep network U-Net and its variants,combined with attention mechanism,and focuses on the application of low-field knee MR image segmentation and model reconstruction,conducting research on two image segmentation methods.(1)A nested encoder-decoder architecture based on the U-Net++ network,combined with attention gate and an enhanced classification-guided module,was proposed.This method has three advantages: First,the method can achieve a trade-off between accuracy and speed by flexibly pruning for different image segmentation applications.Second,the attention gate can guide the network to focus on important spatial regions,enhancing the segmentation performance.Third,the enhanced classification-guided module reduces false positive segmentation caused by noise in the feature map by constructing a simple classification task to identify the presence of target tissue in the input image.Experiment results show that the proposed method outperforms the compared methods in terms of image segmentation performance and has good flexibility.(2)We proposed an improved U-Net network architecture based on a large kernel attention mechanism.The encoder part of the network is replaced by a Res Net RS backbone to enhance the feature extraction performance,while the skip connection part uses a large kernel attention mechanism to model longrange dependencies.This method can effectively overcome errors in segmentation caused by noise,blurring,and artifacts in magnetic resonance images,and also performs well in segmenting small tissues and extremely dark areas.Experimental results show that the proposed method outperforms the comparison methods in image segmentation task.In this paper,we also performed model reconstruction based on the image segmentation results of the second proposed method and conducted electromagnetic simulations in a high-field 3T environment to calculate the local SAR of the constructed model,and the results showed that the local SAR peak of the proposed method has a less error rate compared with the comparison methods.This study can help improve the tissue segmentation performance of lowfield knee magnetic resonance joint images,laying a foundation for accurate construction of knee joint models and improving the estimation accuracy of local SAR,thereby promoting the application of high-field magnetic resonance imaging.