Research On Medical Image Segmentation Methods Based On Deep Learning

Medical image segmentation is an important field in medical image analysis,and it is also a necessary part of computer-aided diagnosis,monitoring,intervention and treatment.Its key task is to segment objects of interest(such as organs or lesions)in medical images.Accurate identification,detailed analysis,reasonable diagnosis,prediction and prevention of diseases are of great significance and value.At present,medical image segmentation is mainly performed manually by doctors.This segmentation method is time-consuming,and the segmentation results have a great relationship with the doctor’s clinical experience.With the research on semi-automatic and fully automatic segmentation algorithms,manual segmentation will gradually be used.segmentation algorithm instead.Therefore,accurate automatic segmentation of medical images is of great significance for clinical auxiliary diagnosis.This article analyzes the characteristics and advantages of traditional segmentation methods and deep learning methods,and designs medical image segmentation algorithms based on different deep learning principles.Experiments and analyses were conducted on the retinal vessel public dataset,the clinical data set of femoral-popliteal artery stent,and the polyp public dataset.The results show the effectiveness and robustness of deep learning-based medical image segmentation methods,and also demonstrate the impact and role of the performance of deep learning algorithms on segmentation tasks.For specific tasks,combining deep learning networks with computer vision methods can further improve the performance of the method.The main research content and innovative work of this article are summarized as follows:(1)Aiming at the interference of segmentation targets in complex muscle,blood vessel,bone and other backgrounds and the lack of highlevel semantic information in the network,a medical image segmentation method based on a multi-branch hybrid attention network is proposed.The method mainly consists of an encoder module with a pre-trained Res Net,a lightweight Pyramid Split Attention(PSA)module,and a novel Multibranch Hybrid Attention(MHA)feature decoder module.We use the PSA module as a bridge to connect the encoder and decoder to obtain richer multi-scale feature maps.The proposed MHA block can recover more high-level semantic information by employing concatenation and summation operations combined with corresponding feature maps.Experimental results on retinal vessels,femoropopliteal artery stents and polyp datasets show that our proposed MHA-Net has better performance.(2)The lack of context information for the network in the previous chapter and the continuous pooling and stride convolution operations of UNet lead to the loss of some spatial information,a medical image segmentation method based on U-shaped contextual residual network,called UCR-Net,is proposed to capture more context and high-level information for medical image segmentation.UCR-Net is an encoderdecoder framework.The feature decoder module contains four newly proposed Context Attention Exploration(CAE)modules,a newly proposed Global and Spatial Attention(GSA)module and four decoder modules.The CAE module captures more multi-scale context features from the encoder.The GSA module further explores global contextual features and semantically enhanced deep features.Experiments on retinal vessels,femoropopliteal artery stents and polyp datasets show that the proposed UCR-Net performs well compared to the original U-Net and other state-ofthe-art methods.(3)In view of the lack of global information in the current network and the diversity and camouflage of medical segmentation targets,a MSDM-Net network based on Transformer model combined with MultiScale Convolutional Attention(MSCA)module and Distraction Mining(DM)module is proposed to achieve effective medical image segmentation.Local and global information can help effectively segment diverse medical segmentation targets.Therefore,we propose a Multi-Scale Convolutional Attention(MSCA)module,the multi-scale convolutional attention weight of the lower layer guides the current layer,which is beneficial to extract multi-scale information and enhance the current features.In addition,the color and texture of medical segmentation objects are very similar to surrounding tissues with low contrast.There is a high degree of similarity between the segmented object and the background,which provides it with strong camouflage properties and makes it difficult to recognize.Therefore,we propose the Distraction Mining(DM)module to gradually refine the coarse prediction results by distraction mining in ambiguous regions to improve the prediction performance.Experiments on retinal vessels,femoropopliteal artery stents,and polyp datasets show that MSDM-Net performs well.