RANet: An end-to-end residual-aware adaptive registration network for echocardiogram

Echocardiographic registration is crucial for cardiac disease screening, but challenges such as low contrast, speckle noise, and motion artifacts complicate the deformable registration task. Current achievements mainly focus on intro-subject left ventricle motion tracking, while facing challenges of registering narrow and vague myocardial wall. To overcome these limitations, we propose a more universal residual-aware adaptive registration network (RANet), exhibiting good performance on main structures like ventricular, myocardium, and atria. This end-to-end framework utilizes parallel adaptive wavelet convolution (WTConv)-convolutional neural network (CNN) fusion in its parallel enhanced encoder (PEE) to extract complementary global-local features at multiple scales, strategically directs attention to misaligned structural edges through residual features in the multi-head residual attention (MRA) module, and integrates channel-space-contextual attention with learnable weights via the adaptive collaborative attention (ACA) module for noise-robust feature fusion and adaptive contribution balancing across hierarchies. Experiments are evaluated on three datasets. The results demonstrate that RANet outperforms state-of-the-art methods in both accuracy and robustness. • We propose end-to-end RANet for echocardiographic deformable registration. • Integrating adaptive wavelet-CNN to expand receptive fields with minimal parameters. • Residual-aware attention pinpoints misaligned areas for registration. • Adaptive Collaborative Attention enables optimized feature transfer and fusion. • RANet is robust and accurate on three datasets in two application scenes.