RECENet: A Dual-Stream Contrast-Aware Attention Network With Bidirectional Pyramid Consistency for CEST Image Registration.
L’essentiel
To develop and evaluate a deep learning framework for deformable registration of CEST MRI that achieves accurate alignment under substantial interoffset contrast variations. A deep learning architecture, termed Registration CEST Network (RECENet), was developed for deformable registration of CEST images. The framework employed a dual-stream multiscale encoder to independently extract hierarchical features from image pairs acquired at different frequency offsets, preserving contrast-specific information while enabling interoffset correspondence learning. A multilevel contrast-aware attention fusion module was integrated at the bottleneck to recalibrate spatial and channel representations, enhancing alignment under heterogeneous contrast conditions. A bidirectional pyramid decoder progressively estimated forward and backward deformation fields in a coarse-to-fine manner, with inverse-consistency constraints to ensure physically plausible deformations. RECENet was evaluated on multiorgan CEST datasets, including brain, lower extremity, and abdominal imaging, acquired on multiple scanners and field strengths under both synthetic and real deformations. Quantitative comparisons with conventional and deep learning-based methods showed that RECENet achieved higher Dice coefficients, local normalized cross-correlation, mutual information, and structural similarity index, as well as lower root mean square error. Improved spatial alignment enhanced Z-spectrum fidelity and increased metabolite quantification accuracy. RECENet provides a robust and generalizable solution for contrast-invariant deformable registration of CEST MRI, improving spatial alignment and quantitative accuracy.
Synthèse détaillée
Résumé original
To develop and evaluate a deep learning framework for deformable registration of CEST MRI that achieves accurate alignment under substantial interoffset contrast variations. A deep learning architecture, termed Registration CEST Network (RECENet), was developed for deformable registration of CEST images. The framework employed a dual-stream multiscale encoder to independently extract hierarchical features from image pairs acquired at different frequency offsets, preserving contrast-specific information while enabling interoffset correspondence learning. A multilevel contrast-aware attention fusion module was integrated at the bottleneck to recalibrate spatial and channel representations, enhancing alignment under heterogeneous contrast conditions. A bidirectional pyramid decoder progressively estimated forward and backward deformation fields in a coarse-to-fine manner, with inverse-consistency constraints to ensure physically plausible deformations. RECENet was evaluated on multiorgan CEST datasets, including brain, lower extremity, and abdominal imaging, acquired on multiple scanners and field strengths under both synthetic and real deformations. Quantitative comparisons with conventional and deep learning-based methods showed that RECENet achieved higher Dice coefficients, local normalized cross-correlation, mutual information, and structural similarity index, as well as lower root mean square error. Improved spatial alignment enhanced Z-spectrum fidelity and increased metabolite quantification accuracy. RECENet provides a robust and generalizable solution for contrast-invariant deformable registration of CEST MRI, improving spatial alignment and quantitative accuracy.