A Symmetric Dynamic Learning Framework for Diffeomorphic Medical Image Registration

TL;DR

This study introduces DCCNN-LSTM-Reg, a learning framework that evolves dynamically and learns a symmetrical registration path by satisfying a specified control increment system to obtain symmetric diffeomorphic deformations between moving and fixed images.

Abstract

Diffeomorphic image registration is crucial for various medical imaging applications because it can preserve the topology of the transformation. This study introduces DCCNN-LSTM-Reg, a learning framework that evolves dynamically and learns a symmetrical registration path by satisfying a specified control increment system. This framework aims to obtain symmetric diffeomorphic deformations between moving and fixed images. To achieve this, we combine deep learning networks with diffeomorphic mathematical mechanisms to create a continuous and dynamic registration architecture, which consists of multiple Symmetric Registration (SR) modules cascaded on five different scales. Specifically, our method first uses two U-nets with shared parameters to extract multiscale feature pyramids from the images. We then develop an SR-module comprising a sequential CNN-LSTM architecture to progressively correct the forward and reverse multiscale deformation fields using control increment learning and the homotopy continuation technique. Through extensive experiments on three 3D registration tasks, we demonstrate that our method outperforms existing approaches in both quantitative and qualitative evaluations.

Figures (7)

• Figure 1: The architecture of symmetric diffeomorphic Cascade CNN-LSTM image registration. (a) DCCNN-LSTM-Reg framework, where the symmetric configuration of the SR module in diffeomorphic image registration comprises two competing pathways that stem from the theory of one-to-one deformation. (b) The U-Net module architecture is designed to capture dual multiscale feature pyramids using shared parameters. The green block represents the multiscale features $\{\bm{F}_X^\ell\}_{\ell=1}^L$ or $\{ \bm{F}_Y^\ell\}_{\ell=1}^L$ extracted from the input images $X$ or $Y$ using the pre-trained U-Net. (c) The symmetrical diffeomorphic image registration module (SR module). A series of $N$ CNN-LSTM blocks are connected in succession to capture the incremental field, which are then gradually integrated to produce the final pair of deformation fields. (d) CNN-LSTM block, where $[\bm{F}_X,\bm{F}_Y]$ represent the features extracted from images $X$ and $Y$ by a pre-trained Unet in Fig.\ref{['fig: overview']}(b), and $[\bm{F}_X\circ\bm{\phi}_{t_{n-1}},\bm{F}_Y]$ are pre-aligned by the previous deformation $\bm{\phi}_{t_{n-1}}$.
• Figure 2: Comparison of DSC scores for each anatomical region between state-of-the-art methodologies and our proposed approach. To enhance clarity, the left and right brain hemispheres were combined into a single region. The structures analyzed included the brain stem (BS), thalamus (Th), cerebellar cortex (CblmC), lateral ventricle (LV), cerebellar white matter (WM), putamen (Pu), caudate (Ca), pallidum (Pa), hippocampus (Hi), 3rd ventricle (3V), 4th ventricle (4V), amygdala (Am), CSF (CSF), and cerebral cortex (CeblC).
• Figure 3: Comparisons with different registration methods for one pair of MRI images. From top to bottom: original images and registered images, local zoom-in of the original images, segmented images, deformation fields, heat maps of Jacobian determinants.
• Figure 4: 3D visualization for one pair of images processed by our DCCNN-LSTM-Reg. From top to bottom: Original image, Segmentation label; from left to right: (a) moving image $X$, (b) fixed image $Y$, (c) results of registration from $X$ to $Y$ ($\bm{X\rightarrow Y}$), (d) results of registration from $Y$ to $X$ ($\bm{X\leftarrow Y}$).
• Figure 5: (a) The feature maps in dual multi-scale feature pyramids, eight 2D slice feature maps are randomly selected from five scales in the two feature pyramids. (b) The feature maps during progressive registration at the last scale, the initial features in DCCNN-LSTM-Reg are gradually deformed during the registration process.