A Research On Segmentation Of Prostate MRI Image Based On Generative Adversarial Networks

Prostate cancer has become a serious disease that jeopardizes the health of older men.The first step in the detection and diagnosis of prostate cancer is to segment the prostate tissue.Magnetic Resonance Imaging has a tendency to improve the clinical examination of prostate disease due to the contrast of the generated images and the fact that the process has no radiation to the human body.However,manually segmenting the prostate MR image will cost the doctor or the relevant experts a lot of energy and time.Therefore,an automatic prostate MRI image segmentation algorithm is needed clinically.Generative adversarial networks(GAN)have been gradually applied to various image processing tasks due to their flexible framework and powerful capabilities,such as target detection,image generation,text-to-image transformation,super-reconstruction,etc.Based on the generative adversarial networks model,we study prostate segmentation.The main research contents are as follows:(1)We studies the generative adversarial networks and their variants,and attempts to apply the generative adversarial networks to the prostate segmentation.We propose a prostate segmentation method based on conditional generative adversarial networks with attention mechanism,using the false region attention networks to pay attention to the generated mask image of prostate.In the generated mask image,the most inaccurate area is found by scoring,and finally the most inaccurate area is corrected by the checker.(2)In order to enable the generator to better capture the features of the prostate,a discriminator is added.The additional discriminator has the same network structure as the original discriminator.The input picture size of the two discriminators is different.The discriminator with small input size has a large receptive field,so that the global feature can be captured,and a globally continuous region is generated;the discriminator with a large input size has a smaller receptive field,so that local features can be captured,causing the generator to generate locally continuous regions.Experiments have shown that after adversarial training,the generator can generate a more accurate prostate region,thus completing the segmentation of the prostate.(3)In the process of generative adversarial networks optimization,network training is often unstable.In order to maintain the stability of network training,the feature matching loss function is taken as part of the whole loss function.When the generated segmented image or the manually segmented image is input to the discriminator along with the prostate MRI,the discriminator extracts the input features to distinguish the two,and the generator wishes to learn such features so that the generated segmented image features are believed as manually segmented.The features of the image are matched as much as possible to "cheat" the discriminator.