Research On Medical Image Segmentation Technology Based On Deep Attention Mechanism

In recent years,the advancement and development of medical imaging technology has been swift,with its applications spanning a broad spectrum,including diagnosis,treatment,and research.Segmentation of medical images is a critical component of image processing,the goal being to separate distinct items in medical images for more precise diagnosis and treatment.With the increasing resolution of images,modern medical images have changed from the traditional two-dimensional images to three-dimensional images,which makes the requirements for medical image segmentation even more stringent and delicate.Widely employed in medical image segmentation,the U-Net network is a deep learning-based method with the benefit of a straightforward network topology and minimal training set data needs.Despite this,U-Net is not without its issues,such as segmentation results being lost at the edges,lengthy training periods,and a single application situation.This paper suggests two attentionbased medical image segmentation algorithms and an intelligent detection system to address these issues.The main work is as follows:(1)The proposed algorithm combines depthwise separable convolution and hybrid attention mechanism to reduce the number of model parameters and operational complexity by dividing the convolution into depth convolution and point-by-point convolution,and replacing the global pooling operation with depth-separable concession with a stride size of 2 to improve the model running speed and accuracy,while improving the generalization ability of the model to a certain extent,and adding the possibility of applying medical image segmentation algorithms on mobile devices.(2)Proposed another algorithm combines the ECA attention module with spatial attention and uses a novel and improved connection strategy to connect the network structures.With this strategy,the network’s segmentation accuracy can be effectively increased,and its need for training data can be decreased.We can effectively increase the variety and size of the training data by using elastic deformation data augmentation methods,which not only significantly improve the dataset’s structure but also significantly enhance the network’s performance.To further enhance its training performance,a local-global training strategy that combines the outcomes of training on local regions as well as training on the entire image is also used.After extensive experiments,it has been found that both of the algorithms suggested in this study can perform significantly better than other algorithms when it comes to segmenting medical images.These improvements include Dice coefficients and IoU metrics.These findings demonstrate the potential and wide range of applications for the algorithms proposed in this paper,which can significantly increase the precision and effectiveness of medical image segmentation.(3)To help patients and doctors analyze and diagnose the stage of the disease,this paper develops an MRI-based Alzheimer intelligent detection system based on the above algorithm.Future studies will investigate in depth how to use this method to solve other medical image processing issues as well as how to raise the level and quality of medical image processing.