Diffeomorphic image registration is a crucial task in medical image analysis. Recent learning-based image registration methods utilize convolutional neural networks (CNN) to learn the spatial transformation between image pairs and achieve a fast inference speed. However, these methods often require a large number of training data to improve their generalization abilities. During the test time, learning-based methods might fail to provide a good registration result, which is likely because of the model overfitting on the training dataset. In this paper, we propose a neural representation of continuous velocity field (NeVF) to describe the deformations across two images. Specifically, this neural velocity field assigns a velocity vector to each point in the space, which has higher flexibility in modeling the complex deformation field. Furthermore, we propose a simple sparse-sampling strategy to reduce the memory consumption for the diffeomorphic registration. The proposed NeVF can also incorporate with a pre-trained learning-based model whose predicted deformation is taken as an initial state for optimization. Extensive experiments conducted on two large-scale 3D MR brain scan datasets demonstrate that our proposed method outperforms the state-of-the-art registration methods by a large margin.
翻译:在医学图像分析中,基于学习的图像登记法是一项关键的任务。最近基于学习的图像登记方法使用进化神经网络学习图像配对之间的空间转换,并达到快速推导速度。然而,这些方法往往需要大量的培训数据以提高其一般化能力。在测试期间,基于学习的方法可能无法提供良好的登记结果,这很可能是由于培训数据集的模型过大所致。在本文中,我们提议对连续速度场(NeVF)进行神经表示,以描述两个图像的变形。具体地说,这个神经速度场为空间的每个点指定了一个速度矢量,在复杂变形场的建模方面具有更大的灵活性。此外,我们提出一个简单的稀有抽样战略,以减少变形登记处的记忆消耗量。拟议的NeVF还可以与一个预先培训的基于学习的模型相结合,该模型的预测变形作为初步优化状态。在两个大型3D MR 脑扫描模型中进行了广泛的实验,通过大容量定位显示我们拟议的方法显示一个大比例的注册方法。