Bohua Chu, Baoju Zhang, Bo Zhang, Cuiping Zhang
CoReg-Net shows that explicitly enforcing structural consistency through contrastive registration improves structural similarity and maintains segmentation accuracy. The proposed framework provides a reliable and generalizable solution for atlas-guided anatomical analysis across heterogeneous imaging modalities. Code is available at https://github.com/RelMono/CoRegNet.
The clinical efficacy of modern medical imaging relies heavily on robust deformable registration for accurate diagnosis and treatment planning. Although deep learning has catalyzed significant progress in this domain, existing techniques encounter substantial hurdles in characterizing complex, non-rigid structural deformations. Addressing these limitations, this study introduces a novel 3D registration framework centered on a dynamic wavelet transform module. By adaptively tuning frequency decomposition filters and employing a frequency-sensitive multi-stream architecture, the proposed model enhances feature representation fidelity. The integration of a multi-scale paradigm further ensures the progressive and precise estimation of displacement fields. Furthermore, we optimize the widely used U-Net registration baseline by introducing a wavelet convolution module tailored to 3D images, thereby enhancing the extraction of global contextual information. To evaluate the model's efficacy, extensive comparisons were conducted using state-of-the-art architectures across four diverse, publicly available brain MRI repositories. The empirical evidence confirms that our method achieves superior registration fidelity, highlighting its potential for practical clinical implementation.