Brain Tumor Image Segmentation Based On Full Convolutional Generative Adversarial Network

As a disease of intracranial tissues,brain tumor is difficult to treat,has a high lethality rate,and poses a serious threat to human health.The rapid development of medical imaging technology has made MRI with non-invasive imaging technology has become the preferred modality for the diagnosis and treatment of brain tumors.The uncertainty of location,diversity of shape and complexity of size of brain tumors,it is difficult for clinical diagnosis and treatment.Currently,brain tumor MRI image segmentation mainly relies on manual annotation by clinicians,which is susceptible to personal experience and leads to errors.Therefore,efficient and accurate segmentation of brain tumor images using MRI is of great practical importance in helping physicians diagnose,develop surgery and post-operative recovery plans.In recent years,deep learning has rapidly developed in the field of computer vision,and many deep learning-based algorithms are being applied to the domain of medical images,which are of important research significance in helping the diagnosis and treatment of brain tumors and other diseases.Problems such as sample class imbalance and model design led to limited accuracy in brain tumor segmentation,the study found.For the above reasons,this thesis utilizes full convolutional network and generative adversarial network to conduct an in-depth study of brain tumor MRI image segmentation task,as follows:Firstly,a brain tumor segmentation method based on full convolutional network(BTSUNet)is proposed,which is based on U-Net and streamlines one layer of network structure to reduce the loss of low-level semantic information of images.Also,to alleviate the sample class imbalance problem,weighted cross-entropy loss and generalized dice loss are combined to reduce the difference between training samples and evaluation metrics,so that the model can be good at learning difficult to classify samples.The contrast experiment shows that this method can effectively improve the segmentation accuracy of brain tumor.Secondly,a method of multimodal brain tumor image segmentation based on a full convolutional network(ARFU-Net)with receptive field and attention mechanism is proposed.The method is based on BTSU-Net,and the perceptual field module with multi-branch convolution and dilated convolution is embedded into the decoder and encoder of the model to increase the perceptual field while fusing multi-scale information,which can better localize the tumor location.In addition,the attention module is embedded in the cascade structure of the decoder part to reduce the redundant information while effectively enhancing the local response of the sampled features.The method not only can effectively analyze the complex distribution of brain tumor images,but also facilitates the extraction of more detailed tumor information.A large number of comparison tests were conducted for the model,and the segmentation accuracies were 88.7%,81.1% and 77.2% for the overall tumor,core tumor and enhanced tumor on the Bra Ts2018 dataset,and 87.8%,79.4% and 72.2% on the Bra Ts2019 dataset,respectively,and the proposed model achieves accurate brain tumor automatic segmentation.Finally,a semantic brain tumor segmentation method(NFGAN)based on generative adversarial networks is proposed.NFGAN is built on GAN taking brain tumor MRI images as model input,and a modified full convolutional network(ARFU-Net)is used as a generator to achieve end-to-end segmentation results.A convolutional neural network(DCN)is also used as a discriminator to increase the adversarial training and further improve the model segmentation performance.the dual network structure of NFGAN is better able to learn the sample data features adversarially and focus on the potential probability density of the data during training.Extensive experiments were conducted for the model,and the segmentation accuracies were 89.2%,82.7% and 78.4% for each category of tumors on the Bra Ts2018 dataset,and 88.1%,80.9% and 73.5% on the Bra Ts2019 dataset,respectively,which showed further improvement of brain tumor segmentation accuracy.