| Accurate segmentation of medical images is of great significance for rapid diagnosis and treatment of diseases.As the most popular semantic segmentation network,U-net has achieved good application effect in the field of medical image segmentation.However,Unet still has some issues such as low precision of small target segmentation and slow convergence in training,and its structure and setting of hyperparameters have great influence on network performance.When faced with a specific segmentation task,it is often necessary to manually design the U-net architecture to obtain satisfactory segmentation results,which is a process that is highly dependent on expertise and is constantly trial-and-error.Therefore,the automatic optimal design of neural network architecture has attracted extensive attention,among which the neural network architecture design method based on evolutionary algorithms is one of the most widely used methods.As a new evolutionary intelligent optimization algorithm,state transition algorithm has the advantages of global,rapidity and controllability.This algorithm has potential advantages in dealing with neural network architecture optimization problems.In this paper,the U-net architecture optimization method based on state transition algorithm is studied.The main contributions of this thesis are summarized as follows.1)Considering the unknown problem of the optimal depth of U-net architecture,a flexible variable depth representation method is designed to better explore the potential optimal architecture of the network.Variable depth means that the number of up-sampling and down-sampling times of U-net network is not limited.Under this variable depth coding scheme,the evolving architecture is expected to achieve good segmentation performance when solving different tasks.2)A U-net architecture optimization method based on discrete state transition algorithm is proposed to solve the combinatorial explosion problem in network search space.Hyperparameters in U-net architecture are characterized by high dimension,discreteness and interdependence.It is very difficult to find the optimal network architecture by manually adjusting hyperparameters.Therefore,a two-layer coding strategy is adopted to represent individuals of different network architectures flexibly.In order to speed up the evolution of network,transfer learning strategy is used to transfer the optimal genes learned in the sub-data set to the original big data set for training.The effectiveness of the algorithm is verified by experiments on two data sets with different image characteristics.3)To solve the problem of serious foreground background imbalance in small target images,an attention U-net model combined with improved focus loss function is proposed.The focus loss function can solve the problem of serious imbalance between positive and negative samples,and the introduction of attention module can help extract deeper detail features.In order to obtain the optimal architecture of the network,the mixed state transfer algorithm is used to further optimize the network,and the validity of the algorithm is verified by the segmentation experiment of the inner and outer membrane of the left ventricle.Finally,an automatic segmentation system of left ventricular images based on U-net architecture optimization is developed,which has a certain guiding role in assisting doctors in medical diagnosis and treatment.There are 40 figures,11 tables,99 references. |
