ARCNet: Adaptive Reconstruction-Driven Cascaded Network for Deformable Registration of Images With Pathologies

Abstract

Deformable registration is a critical task in medical image analysis. However, registration of images with tumors is challenging due to absent correspondences induced by the tumor. Furthermore, disease progression or normal aging may cause more intricate deformations in the brain. Therefore, this paper proposes a new adaptive reconstruction-driven cascaded network (ARCNet). Specifically, the symmetric-constrained feature reasoning (SFR) module is designed to reconstruct tumor regions without valid correspondence as normal tissue, allowing the establishment of dense correspondences during the registration process. The dilated multi-receptive feature fusion (DMFF) module is further introduced, which collects long-range features from different dimensions and helps generate well-structured content in the tumor region reconstruction, especially for large tumor cases. Then an adaptive importance-aware guidance module (AIG) is proposed, which adjusts the local importance of a region according to the deformation complexity, directing the network to focus on difficult-to-align regions with complex deformations, thus improving the registration accuracy. We conducted experiments on the BraTS 2021 dataset to validate the effectiveness of the SFR, DMFF, and AIG modules. Using quantitative metrics such as Dice Similarity Coefficient (Dice), the Local Normalized Cross-Correlation (LNCC), the negative Jacobian determinant percentage (%|J| ≤ 0), the 95% Hausdorff Distance (HD95), and Average Surface Distance (ASD), experimental results show that the proposed method effectively handles the problem of pathological image registration, which can maintain the smooth deformation of the tumor region while maximizing the image similarity of normal regions.