在基于学习的医疗图像注册中,以网格控制点的自由度为准
Wen Yan1, Qianye Yang1, Yipei Wang1
1Department of Medical Physics and Biomedical Engineering, UCL Hawkes Institute, University College London, London, UK.
Medical physics
|February 19, 2026
概括
这项研究介绍了GridReg,这是一种用于医疗图像注册的新方法,它使用稀疏的控制点而不是密集的场. GridReg显著提高了计算效率和注册准确性,提供了更稳定,更紧的变形表示.
科学领域:
- 医学图像分析 医学图像分析
- 计算解剖学的计算解剖学
- 机器学习在医学成像中的应用
背景情况:
- 沃克塞尔智能注册方法面临着高维度和在同质或噪音图像区域中的不良问题.
- 在注册网络中密集的voxel-wise解码器可能导致过度的内存使用和稳定性降低.
- 稀疏的控制点参数化为变形提供了紧而光滑的表示,提高了稳定性并减少了内存足迹.
研究的目的:
- 调查开发基于学习的医学图像注册网络所需的最佳控制点数量.
- 为了比较稀疏的控制点配置 (例如,5x5x5) 与分散的控制点和密集的位移场等替代方法.
主要方法:
- 引入了基于学习的框架GridReg,它用在稀疏控制点上的位移预测取代了密集的voxel-wise解码.
- 采用了具有位置编码的多尺度3D编码器,将特征地图平成一个1D令牌序列,以保持空间上下文.
- 利用交叉注意模块来预测稀疏的网格变形场,其中控制点通过关注本地编码器令牌来估计位移,然后插入到密集的场.
- 实施了网格适应训练,使模型在推断过程中能够适应多个网格大小,而无需重新训练.
主要成果:
- 通过预测稀疏电网样本位移,在计算效率方面取得了显著的改进.
- 与VoxelMorph,TransMorph和KeyMorph等现有方法相比,在可比或降低的计算成本下实现了优异的注册性能.
- 通过使用三个不同的数据集,在包括前列腺,骨盆器官和神经组织在内的各种解剖结构中验证了益处.
结论:
- 预测稀疏网格位移可以降低计算成本和/或提高医疗图像注册的性能,无论编码器架构如何.
- GridReg方法很容易实现,可以适应各种需要灵活的网格大小的注册任务.
- 开发的框架和代码是公开可用的,用于进一步的医学图像注册的研究和应用.
相关概念视频
Degrees of Freedom
10.4K
The degree of freedom for a particular statistical calculation is the number of values that are free to vary. As a result, the minimum number of independent numbers can specify a particular statistic. The degrees of freedom differ greatly depending on known and uncalculated statistical components.
For example, suppose there are three unknown numbers whose mean is 10; although we can freely assign values to the first and second numbers, the value of the last number can not be arbitrarily...
For example, suppose there are three unknown numbers whose mean is 10; although we can freely assign values to the first and second numbers, the value of the last number can not be arbitrarily...
10.4K
Degrees of Freedom
7.3K
The degree of freedom for a particular statistical calculation is the number of values that are free to vary. Thus, the minimum number of independent numbers can specify a particular statistic. The degrees of freedom differ greatly depending on known and uncalculated statistical components.
For example, suppose there are three unknown numbers whose mean is 10; although we can freely assign values to the first and second numbers, the value of the last number can not be arbitrarily assigned.
For example, suppose there are three unknown numbers whose mean is 10; although we can freely assign values to the first and second numbers, the value of the last number can not be arbitrarily assigned.
7.3K


