Research On Multi-scale Denoising Enhanced Deep Learning Segmentation Network Construction And Application Of Multi-dataset Medical Images

Background and ObjectiveThe recognition,detection,classification and segmentation of medical lesions play an important role in clinical diagnosis and treatment.Colorectal cancer(CRC),for example,is the third most common cancer after breast cancer and lung cancer.In order to reduce the endoscopic error rate of colorectal polyps and assist clinical endoscopists to locate complex colonic polyps,computer-aided polyp segmentation system is particularly important.Compared with traditional segmentation methods,medical segmentation methods using deep learning have stronger learning and adaptability,and show strong advantages in overall operation efficiency and accuracy of segmentation.As a reference network for medical image segmentation,UNet consists of encoder,decoder and jump connection structure.However,in the process of direct fusion of low-level and high-level convolutional features with semantic gaps,traditional jump connections may cause problems such as fuzzy feature maps and segmentation errors of target regions.In order to deal with the problem of low segmentation accuracy caused by this kind of situation,this paper aims to improve the jump connection innovatively.Specifically,we propose an improved jump-connection structure,which can better handle the information fusion between different levels and scales of features,thus improving the segmentation accuracy.In addition,traditional medical segmentation methods are mostly trained and validated on limited data sets,and are not extended to more medical segmentation data sets to complete the universal validation of segmentation methods.In order to solve this problem,this study innovatively introduces the concept of universal segmentation of medical image multiple datasets.The effectiveness and generality of the proposed method are verified on several medical segmentation datasets.Contents and MethodsThe MDE module proposed in this paper is a universal partition module,which can be applied to different network backbones.In order to verify its versatility,this paper has carried on the research in the different network backbone,and has carried on the verification experiment separately.1.in this pap,that traditional skip concatenation is replaced by a multi-scale denoise enhancement(MDE)module before the encode and decoder features are merged.Before fusing the encoder and decoder feature maps,the encoder feature map is first processed using depth convolution with spatial enhancement filter to generate an enhanced intermediate feature map and fused with the decoder feature map.2.We try to integrate the MDE module with the attention family in order to achieve better results.The attention family includes Attention-UNet and SEgmentation Transformer(SETR).3.We try to fuse the MDE module with the UNet family in order to achieve better results.The UNet family includes UNet,Res-UNet,Dense-UNet.MDE is applied to multi-modality medical image segmentation in order to improve the segmentation performance and efficiency.The MDE module is a new type of segmentation module,which can better fuse the features with semantic differences and enhance the feature graph of the whole network layer,so as to improve the performance and generalization ability of the network.Through the above innovative improvements to the skip connection,the problem of low segmentation accuracy caused by such situations can be dealt with in a targeted manner.ResultsTraining is performed on polyp(colonoscopy)segmentation dataset,and generalization validation is performed on a total of 6 datasets including lung(CT,X-ray),brain(MR),cell(cytoscope),skin(cytoscope)and fundus(angiography),which have different medical modalities.On the Attention family and the UNet family,we combined MDE on different baseline networks to perform ablation studies with DSC and F2 values as the primary measures,taking polyps(colonoscopy)as an example,where the DCS value of the best model MDE-Transformer in the Attention family was 96.47%,and the DSC value of the best model E-DU in the UNet family was 98.85%.Evaluation and comparison of segmentation performance showed that:E-DU achieved the most excellent segmentation results in DSC and other evaluation indexes,with DSC values(%)of 98.85,97.64,95.31,94.42,94.93,97.78 and 98.38,respectively.The F2values(%)were 98.64,97.73,93.50,95.20,96.64,97.39 and 98.51,respectively.The segmentation efficiency results show that E-DU has the fastest convergence speed,and the addition of MDE module can effectively save 0.1-0.2G of peak memory usage(PMU)and reduce the average time per step(ATS)by about 70-90 seconds.Polyp(colonoscopy)is used as an example to illustrate the improvement of segmentation performance.In the attention family,the performance of Attention network and Transformer network improved by1.44 and 1.43(DSC/%)respectively after adding MDE module.In the UNet family,the performance of the UNet,Res UNet and Dense UNet networks with the MDE module is improved by 4.75,2.38 and 2.26(DSC/%)respectively.Compared with before,the addition of MDE module can achieve better segmentation effect,and is effectively proved by visual prediction map,which further illustrates its versatility and high segmentation accuracy.ConclusionThe E-DU in this paper has good segmentation effect and running efficiency,and can be flexibly extended to multi-modality medical segmentation data sets.The experimental results show that the E-DU model achieves good segmentation results on different data sets and has high versatility.The E-DU model has a fast running speed and can complete the image segmentation task in a short time,so it has a high practical value.The E-DU model uses several dense connection structures,and realizes the effective processing of multi-scale images through the techniques of jump connection and multi-scale feature fusion.At the same time,E-DU model also introduces noise reduction and enhancement module,which can effectively improve the performance and generalization ability of the model.The idea and method of this paper can realize clinical multi-scale image segmentation,and then make full use of image information to provide clinical decision support,which has great application value and promotion prospect.In the field of medical image analysis,image segmentation is a very important task,which can provide important clinical decision support for doctors.The proposed E-DU model can effectively process multi-scale images and fuse semantically different features,providing more excellent tools and methods for medical image analysis.