Attention-enhanced Dual-stream Registration Network via Mixed Attention Transformer and Gated Adaptive Fusion

Abstract

Deformable registration requires extracting salient features within each image and finding feature pairs with potential matching possibilities between the moving and fixed images, thereby estimating the deformation field used to align the images to be registered. With the development of deep learning, various deformable registration networks utilizing advanced architectures such as CNNs or Transformers have been proposed, showing excellent registration performance. However, existing works fail to effectively achieve both feature extraction within images and feature matching between images simultaneously. In this paper, we propose a novel Attention-enhanced Dual-stream Registration Network (ADRNet) for deformable brain MRI registration. First, we use parallel CNN modules to extract shallow features from the moving and fixed images separately. Then, we propose a Mixed Attention Transformer (MAT) module with self-attention, cross-attention, and local attention to model self-correlation and cross-correlation to find features for matching. Finally, we improve skip connections, a key component of U-shape networks ignored by existing methods. We propose a Gated Adaptive Fusion (GAF) module with a gate mechanism, using decoding features to control the encoding features transmitted through skip connections, to better integrate encoder–decoder features, thereby obtaining matching features with more accurate one-to-one correspondence. The extensive and comprehensive experiments on three public brain MRI datasets demonstrate that our method achieves state-of-the-art registration performance.