MCAM-Net: A non-iterative multi-scale coarse-to-fine registration network combined with adaptive morphology for deformable image registration

Abstract

Deformable image registration has been a major focus of medical image analysis. However, deep-learning-based deformable registration methods may be limited by the analysis ability to the complex textural features and the ability of establishing matching relationships in regions of difference between image pairs. Moreover, when morphology is employed for registration, the morphological operation’s flexibility may limit the effectiveness of morphological information for improving registration performance. This paper proposes MCAM-Net, a non-iterative multi-scale coarse-to-fine registration network combined with adaptive morphology for deformable image registration. MCAM-Net consists of a multi-scale coarse-to-fine registration network (MCR-Net) and an adaptive morphology network (AM-Net). MCR-Net employs a multi-scale feature learning (MFL) encoder to learn multi-scale features, enhancing its ability to analyze the complex textural features. It also utilizes the differential region guidance (DRG) block to guide the network in establishing matching relationships in regions of difference between image pairs. AM-Net can adaptively obtain the adaptive local-signed-distance field (ALSDF) without manually adjusting the dilation and erosion intensity, improving registration performance. Experimental results from three publicly available brain MR datasets and one liver CT dataset demonstrate that MCAM-Net outperforms state-of-the-art medical image registration methods. Ablation studies show that the proposed model components are effective for deformable image registration.