一种改进的小样本医学图像分割算法研究
Research on an improved few-shot medical image segmentation algorithm
罗兆林 1宋亚男 1徐荣华 1萧飞鹏1
作者信息
- 1. 广东工业大学,广东 广州 510006
- 折叠
摘要
文章针对医学图像的小样本特点和分割模型泛化能力差的问题,提出了一种基于元学习的医学图像分割网络.文章首先在3D U-Net网络基础上,将其下采样模块从两层三维卷积层增加为三层,将每层三维卷积层的批归一化改进为组归一化;其次在U型网络编解码器连接处,引入Transformer模块,增强模型提取全局信息的能力;在U型网络跳跃连接处引入了改进的注意力门机制,原理是替换其中的批归一化改进为组归一化,优化使用低批次来训练模型的效果,将Softmax激活函数替换为ReLU激活函数;最后使用模型无关元学习(Model Agnostic Meta Learning,MAML)算法训练模型.在公开数据集 M&Ms 上的实验结果表明,文章算法的 Dice 评分和 Hausdorff 距离分别为 69.9%和11.88 mm,与其他主流算法对比,分割精度更优,泛化能力更好.
Abstract
In view of the small sample characteristics of medical images and the poor generalization ability of segmentation models,the paper proposes a medical image segmentation network based on meta-learning.Firstly,based on the 3D U-Net network,the down-sampling module is increased from two layers of 3D convolutional layers to three layers,and the batch normalization of each layer of 3D convolutional layers is improved to group normalization.Secondly,at the connection of the U-shaped network codec,a Transformer module is introduced to enhance the model's ability to extract global information.An improved attention gate mechanism is introduced at the U-shaped network skip connection,which replaces batch normalization with group normalization,optimizes the effect of using low batches to train the model,and replaces Softmax activation function with ReLU activation function.Finally,the model is trained using Model-Agnostic Meta-Learning(MAML)algorithm.The experimental results on the public dataset M&Ms show that the Dice score and Hausdorff distance are 69.9%and 11.88 mm.Compared with other mainstream algorithms,the model segmentation accuracy and generalization ability is better.
关键词
小样本学习/元学习/医学图像分割/注意力机制Key words
few-shot learning/meta-learning/medical image segmentation/attention mechanis引用本文复制引用
基金项目
广东工业大学高水平大学建设研究生教育创新计划(2018JGMS-09)
广东省本科高等学校在线开放课程指导委员会研究课题(2022ZXKC143)
出版年
2024
NETL
NSTL
万方数据