LDRNet: Large Deformation Registration Model for Chest CT Registration

TL;DR

LDRNet addresses large-deformation chest CT registration by a fast unsupervised coarse-to-fine network that integrates a refine block and a rigid block to capture local and global alignment. The method jointly refines a deformation field across resolutions while applying a learned rigid transform to guide global matching, guided by a loss combining $L_{sim}$ and regularizers $L_{range}$ and $L_{smooth}$ with explicit bounds on the field. Across SegTHOR and private chest CT datasets, LDRNet surpasses common baselines such as VoxelMorph, ANTs, and LapIRN in accuracy and is significantly faster, enabling real-time-like performance. The results are supported by extensive ablations and qualitative visualizations, though regional continuity and generalization with limited data remain areas for improvement.”

Abstract

Most of the deep learning based medical image registration algorithms focus on brain image registration tasks.Compared with brain registration, the chest CT registration has larger deformation, more complex background and region over-lap. In this paper, we propose a fast unsupervised deep learning method, LDRNet, for large deformation image registration of chest CT images. We first predict a coarse resolution registration field, then refine it from coarse to fine. We propose two innovative technical components: 1) a refine block that is used to refine the registration field in different resolutions, 2) a rigid block that is used to learn transformation matrix from high-level features. We train and evaluate our model on the private dataset and public dataset SegTHOR. We compare our performance with state-of-the-art traditional registration methods as well as deep learning registration models VoxelMorph, RCN, and LapIRN. The results demonstrate that our model achieves state-of-the-art performance for large deformation images registration and is much faster.

Figures (5)

• Figure 1: Difference between large deformation chest images (up) and small deformation brain images (down)
• Figure 2: Overview of our proposed registration model. It includes a feature extraction path(blue), a refine path(green), a pooling path(orange) and a rigid block. The number of channels $C_0, C_1, C_2, C_3$ in the feature extraction path are 8, 16, 32, 64 respectively
• Figure 3: Refine block. The input includes four components, feature from feature extraction path, fixed image and moving image from pooling path, and registration field in smaller resolution from previous stage. The area enclosed by the blue box is $Refine \ Core$, also called as $RC$.
• Figure 4: Images of different resolutions from coarse(left) to fine(right) and registration fields generated by each refine block.
• Figure 5: Experiments of different combination of $\alpha$(x-axis) and $\beta$(y-axis). x-axis is $\log \alpha$. y-axis is $\log \beta$. z-axis is dice scores.