基于状态转移算法和U-net的医学图像分割

Medical image segmentation based on state transition algorithm and U-net

U-net在医学图像分割领域应用广泛,但存在小目标分割精度低、模型收敛慢等问题,且其结构和超参数的设定对网络性能有很大影响。为此,本文提出基于混合状态转移算法的U-net结构设计方法,以获取不同分割任务下的较优的U-net体系结构。首先,提出一种可变深度的编码策略来表示U-net中不同的构建块和潜在的最优深度;其次,通过混合状态转移算法优化网络结构中的超参数和连接权重初始值;再次,设计一种新的交互操作来生成具有潜力的个体,利用迁移学习策略和减少epoch的方法加速网络个体的进化;最后,在心脏MRI、肝脏LiTS这2个医学图像数据集中进行测试,验证本文方法的有效性。研究结果表明:与经典的语义分割网络相比,本文所提方法在Dice、Jaccard、VOE等分割性能评价指标中有更好的表现,验证了本文所提算法的可行性和有效性。

U-net has been well applied in medical image segmentation,but there exist some problems such as low segmentation accuracy and slow model convergence for small targets,and its network performance is greatly affected by structures and hyperparameters.Therefore,a U-net architecture design method based on hybrid state transition algorithm was proposed to achieve a better U-net architecture when facing different medical image segmentation tasks.Firstly,a variable-depth encoding strategy was proposed to represent the different building blocks and potential optimal depths in U-net.Then,the hyperparameters and initial connection weights in the network structure were optimized by hybrid state transition algorithm.In addition,a new interactive operation was developed to generate potential individuals,and the transfer learning strategies as well as the method of reducing the number of epochs were used to accelerate the evolution of networks.Finally,experimental studies were carried out on cardiac MRI and LiTS medical image datasets to verify the effecitveness of the proposed method.The results show that compared with the classical semantic segmentation methods,the proposed method has better segmentation performance in Dice,Jaccard,VOE and other evaluation indexes,which has demonstrated the feasibility and effectiveness of the proposed method.