Research On Urine Sediment Microscopic Image Synthesis Based On Generative Adversarial Network

The main components in urine come from the excrement produced after the blood is filtered and reabsorbed by the kidneys,which are closely related to the blood circulation,urinary system,and circulatory system of the human body.Therefore,urine testing has important value for medical research and clinical diagnosis.Due to the urine sediment detection technology is restricted by the poor stability of microscopic imaging,the current urine sediment detection method still relies on manual microscopic examination.Clinically,urine sediment images collected through microscope imaging are prone to be out of focus,which leads to excessive reliance on subjective empirical judgment in the process of manual microscopic examination.The research in this paper aims to solve the problems of the overall defocusing of the urine sediment microscopic image and the blurred edge of the formed components.Our works are to avoid the influence of subjective factors on the urine sediment detection results,and further promote the standardization and automation of urine sediment detection technology.According to mainstream adversarial generative network algorithms,this paper proposes an image synthesis method based on cascaded conditional generative confrontation network and self-attention based image synthesis method.The main work is as follows:(1)According to the characteristics of the data structure of the image sequence,this paper proposes a urinary sediment microscopic image synthesis method based on the cascade condition generation adversarial network.On the basis of general cascade,we add the crossconnection between adjacent generators and combine the local constraint discriminator to improve the generalization ability and robustness of the model.In addition,this paper extends Re LU as the 2D-Re LU activation unit in the 2D space and incorporates it into the residual learning modules.Through comparison experiments with current mainstream methods and model ablation experiments,the superiority of the cascade network in the task of urinary sediment microscopic image synthesis and the effectiveness of related innovative and improved measures are proved.(2)To suppress the distortion in the synthesized images,this paper combines the selfattention mechanism on the basis of cascade network.In order to prevent instance normalization from weakening the semantic layout information in the feature map,this paper proposes a self-attention instance normalization(SAIN),which is applied to the residual learning module.In regard to the complex red blood cell adhesion scene,this paper introduces the red blood cell counting network in part of the discriminator and reconstructs loss functions.The experimental results indicate that the SAIN and the red blood cell counting network can effectively compensate for the defects of the cascade network and improve the quality of the composite image.In addition,our method is superior to existing methods on indicators...