Brain Deformable Registration Using Global and Local Label-Driven Deep Regression Learning in the First Year of Life

Abstract

Accurate medical image registration is highly important for the quantitative analysis of infant brain dynamic development in the first year of life. However, the deformable registration of infant brain magnetic resonance (MR) images is highly challenging for the following two reasons: First, there are very large anatomical and appearance variations in these longitudinal images; Second, there is a one-to-many correspondence in appearance between global anatomical tissues and the small local tissues therein. In this paper, we use a CNN (convolution neural network)-based global-and-local-label-driven deformable registration scheme. Two to-be-registered image patches are input into the UNet-style regression network. Then, a dense displacement field (DDF) between them is obtained by optimizing the total loss function between two corresponding label patches. Global and local label patches are used only during training. During inference, two new MR images are divided into many patch pairs and fed into the trained network. By averaging the deformation of the patches at the same location, the final 3D DDF between the two whole images is obtained. The highlight is that the global (white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF)) and local tissues can be registered simultaneously without any prior ground-truth deformation. Especially for the local hippocampal tissues, the Dice ratios are substantially improved after registration via our method. Experimental results are presented on the intrasubject and intersubject registration of infant brain MR images between different time points, and the intersubject registration of brain T1-weighted MR images on the OASIS-1 dataset, according to which the proposed method realizes higher accuracy on both global and local tissues compared with state-of-the-art registration methods.