Research On Medical Image Segmentation Algorithm Based On Improved Adversarial Network

With the rapid development of medical imaging,the detection efficiency and diagnosis accuracy of different diseases have been effectively improved.Among them,the segmentation result of the medical image lesion area is an important basis for doctors as a clinical diagnosis.Therefore,lots of medical image segmentation technologies have emerged,which greatly help doctors improve the efficiency and accuracy of medical diagnosis.However,medical images have problems such as high complexity,uneven noise distribution,and large differences in different medical research parts,which cannot be well resolved by traditional image segmentation methods.In recent years,due to the rapid growth of data scale and the rapid increase in computing power,deep learning methods have shown great potential in agriculture,aerospace,military and other fields.At the same time,deep learning has become more widely used in the field of medical image segmentation and has become a mainstream analysis technology.Compared with traditional image segmentation methods,deep learning methods can independently learn network parameters and extract richer semantic feature.These advantages have greatly improved the images segmentation accuracy.However,large-scale medical image datasets are currently scarce,and the datasets have many problems,such as imbalanced samples.This is a huge challenge for deep learning that requires a large amount of data to train network parameters.In addition,medical image data is often accompanied by the following problems: 1)the presence of artifact interference;2)the boundary of the lesion area is rough and fuzzy;3)the surrounding dense glandular tissue interference,etc.These problems affect the learning ability of the neural network,and reduce the segmentation performance of the network model.Therefore,this paper proposes an improved adversarial network(AMMSP-c GAN)model based on attention mechanism and multi-scale spatial pooling,which can effectively realize automatic segmentation of medical images.The model contains two structures: a generator and a discriminator.The generator adopts a U-shaped network with Attention Mechanism(AM)and the discriminator adopts a classic CNN model with a Multi-scale Spatial Pooling(MSP)module,which selects conditional generative adversarial network(c GAN)framework for the whole training process.The segmentation results of medical parts of human breast masses and retinal fundus images demonstrate the segmentation performance and cross-domain applicability of AM-MSP-c GAN model.This article mainly conducts the following research work:(1)This work introduces the related background and progress of medical image and the use of computer-aided diagnosis technology for the image segmentation,mainly including the development of medical image segmentation based on traditional methods and deep learning methods.(2)Respectively introduce and analyze the deep learning model based on the encoderdecoder structure and the adversarial learning method.At the same time,analyze and summarize its shortcomings and related improvement measures.(3)This paper proposes the AM-MSP-c GAN model and introduces the related principles of the model in detail.Among them,the introduction of the AM module into the generator can better capture the feature map of the target area,while suppressing the feature response of the irrelevant area,thereby improving the segmentation result.At the same time,the MSP module is introduced into the discriminator to make it have multi-scale semantic features,thereby improving the discrimination performance.In addition,we use adversarial learning to train the proposed model to comprehensively improve the segmentation ability.(4)This article compares the proposed model with other deep learning models through different experiments.Comprehensive experimental results show that the proposed model achieves higher segmentation accuracy on different evaluation indicators.In addition,through self-ablation experiments,it is proved that the different modules in the proposed model all contribute to the segmentation result,which also verifies the reliability and effectiveness of the method from another perspective.(5)In the experiment,we select four different medical image datasets for qualitative and quantitative analysis.For the breast mass experiment,the CBIS-DDSM and INbreast datasets are used.The INbreast dataset needs to be obtained by sending an email to the relevant agency and signing a statement of responsibility.For the retinal fundus image experiment,the DRIVE and STARE data sets are used.This paper introduces and discusses the above four datasets and their segmentation results respectively.(6)This paper uses Dice similarity coefficient(Dice),Sensitivity,Specificity and Hausdorff Distance(HD)as evaluation indicators to quantitatively analyze the experimental results.At the same time,other evaluation criteria(such as box plots and loss curves)are added for qualitative analysis of the experiment.Through quantitative and qualitative analysis,we prove that the proposed method achieves higher segmentation standards and is more suitable for medical image segmentation tasks.