Retinal Image Generation And Segmentation Based On Neural Network Architecture Search

There is a wide variety of fundus lesions that can damage visual function.Segmentation models based on deep learning can help physicians make rapid diagnoses.However,the performance of the models is very dependent on the quality and quantity of the dataset,and the improvement of the architecture requires a lot of expertise.Generating adversarial networks can solve the challenge of the small number of medical image samples and improve the performance of segmentation models by generating images.Neural network architecture search can improve the efficiency of model improvement,and the automatic design of the architecture can be accomplished by some search strategy.This paper proposes a fast and efficient method for generating and segmenting retinal images by using neural network architecture search to construct a generative adversarial network and a two-level nested segmentation network to address the problems of small medical image dataset samples and low automation of architecture improvement.The main work of this paper includes the following three aspects.(1)Improvement of the generative adversarial network based on the gradient descent search strategy.In this paper,three operation sets are designed according to the requirement of implementing different functions at different locations of the codec structure of the generator.By integrating all the operation sets into the corresponding position architectures,a continuous hyperparametric network is constituted.This paper also proposes a split-supervised module to iteratively optimize and update the architecture and model parameters using a gradient descent search strategy,which finally constitutes a stably trained generative adversarial network.(2)The segmentation network is improved based on the evolutionary algorithm search strategy.In this paper,a concise operation set is designed based on a two-level nested U-shaped architecture.A new network coding scheme is proposed to digitally encode the operation set and the network architecture,and the model optimization is completed under the action of the evolutionary algorithm search strategy and the parameter sharing acceleration strategy.A two-level nested U-shaped architecture segmentation network with fewer parameters is finally constituted.(3)Verification of the performance of the model designed by neural network architecture search.The constructed generative adversarial network and segmentation network are validated using the REFUGE challenge and GAMMA challenge in retinal images,respectively.When the designed generative adversarial network generates 400 images,trained with the original REFUGE Challenge dataset,it performs best among the manually designed segmentation networks with a DICE of about 0.9264,an IOU of about0.8673,and an SPC of about 0.9993.When the number of U-shaped architecture stacks is 4,the performance of the designed segmentation network in the GAMMA Challenge dataset exceeds the ranking champion solution,and the average DICE is about 0.9211.The experimental results show that the neural architecture search framework proposed in this paper has good model design capability.This paper has 22 figures,14 tables,and 126 references.