Research Of Medical Image Segmentation And Classification Algorithm Based On Generative Adversarial Network

With the thriving development of image analysis technology and its wide application in various fields,the image segmentation and classification tasks as well as their clinical applications based on medical imaging have been the focus of researchers.Current medical imaging can present the organization and the anatomical structure as well as the form characteristics of the lesion area,following by lesion localization,organ tissue segmentation,feature information extraction techniques to further analyze patient's condition,thus supporting comprehensive information to the radiologists from several dimensions for the disease diagnosis.Scientific and proper application of medical imaging is capable of assisting radiologists and contributes to a more suitable disease diagnosis scheme.However,due to the multiple modalities and huge quantities of medical imaging,manual processing brings burdens to radiologists and contains certain subjectivity,which could result in misdiagnosis or missed diagnosis.As the computer science and artificial intelligence developing,an interdisciplinary medical image processing technology emerged,which could efficiently process large-scale images and improve the objectivity and systematisms of image diagnosis.In this paper,two key problems in pancreas segmentation and glioblastoma multiform(GBM)image classification are studied,and several segmentation and classification models based on deep learning are proposed.Then,the proposed models are respectively applied to pancreas and GBM datasets to achieve pancreas segmentation and pseudoprogression(Ps P)and true tumor progression(TTP)of GBM image classification.The main contents are summarized as follows:(1)An adversarial U-Net(UDGAN)is designed for pancreas segmentation,where the adversarial learning is introduced into a conventional U-Net to prompt the predicted probability maps to resemble the original images owing to the capacity of a generative adversarial network(GAN)to capture data distributions.Furthermore,Multi-Scale Field Selection Module(MSFS)and Multi-Channel Fusion Module(MCFM)are introduced to search for details from multiple levels for segmentation to improve the network performance.The proposed models are evaluated on the NIH pancreas dataset.UDGAN achieves a Dice Similarity Coefficient(DSC)score of 81.09%,which is 3.36%higher than the baseline U-Net with a DSC of 77.73%.The successive involvement of MSFS and MCFM respectively improve the baseline DSC value by 3.40% and 5.10%.Experimental results show that UDGAN is capable of achieving automatic pancreas segmentation,and the application of MSFS and MCFM helps improve the segmentation performance.(2)A novel segmentation algorithm that imposes two-tier constraints on a conventional network through adversarial learning,termed UDCGAN,is proposed for the challenging pancreas segmentation task.This algorithm is equivalent to involving adversarial learning into a segmentation network that has been trained in an adversarial manner.Two-tier constraints further refine the energy functions of the segmentation network,thus effectively contributing to the preservation of detailed information for segmentation.Then,a Multilevel Cue Collection Module(MCCM)is introduced to obtain a novel Dual Adversarial Convolutional Network with Multilevel Cues(DACNMC),where the MCCM contributes to collecting more useful details from different layers for pancreas segmentation.UDCGAN achieves a DSC score of 82.51% on the NIH pancreas dataset,and the DACN-MC obtains a DSC score of 83.93%.Experimental results show that UDCGAN and DACN-MC perform competitive results in the field of pancreas segmentation.(3)An automatic segmentation model using double adversarial networks with a pyramidal pooling module,namely DAN?P,is proposed with the purpose of collecting cues from different receptive fields for segmentation.Furthermore,an AttentionGuided Duplex Adversarial U-Net(ADAU-Net)is designed to enhance the interdependency among pixels and capture contextual details for a more contextualized prediction by involving attention blocks.DAN?P and ADAU-Net respectively achieve DSC scores of 83.31% and 83.76% on the NIH pancreas dataset.Experimental results show that DAN?P and ADAU-Net could effectively handle the challenging task of pancreatic segmentation.(4)A GBM classification model DC-AL GAN based on DCGAN and Alex Net is introduced to distinguish the Ps P from the TTP images.Alex Net is used as discriminator(i.e.,feature extractor)in this work to collect cues,the antagonism and competition relationship in a GAN helps the discriminator extract subtle features.The proposed algorithm is verified on a GBM dataset collected by the Wake Forest School of Medicine in USA.DC-AL GAN achieves accuracy of 0.920 and AUC of 0.947,which is superior to other state-of-the-art Ps P and TTP classification methods.