RECENet: A Dual-Stream Contrast-Aware Attention Network With Bidirectional Pyramid Consistency for CEST Image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Chemical Exchange Saturation Transfer (CEST) MRI is a valuable technique for non-invasive imaging.
• Accurate deformable registration is crucial for analyzing CEST MRI data, especially with inter-offset contrast variations.
• Existing registration methods struggle with the heterogeneous contrast inherent in CEST imaging.

Purpose of the Study:

• To develop and evaluate a deep learning framework for deformable registration of CEST MRI.
• To achieve accurate image alignment under substantial inter-offset contrast variations.
• To enhance spatial alignment and quantitative accuracy in CEST MRI.

Main Methods:

• A novel deep learning architecture, Registration CEST Network (RECENet), was developed.
• RECENet utilizes a dual-stream multiscale encoder for hierarchical feature extraction from different frequency offsets.
• A contrast-aware attention fusion module and bidirectional pyramid decoder with inverse-consistency constraints were employed for robust registration.

Main Results:

• RECENet demonstrated superior performance across multiorgan CEST datasets (brain, lower extremity, abdomen) on multiple scanners and field strengths.
• Quantitative comparisons showed RECENet outperformed conventional and other deep learning methods in alignment accuracy (higher Dice, NCC, MI, SSIM; lower RMSE).
• Improved spatial alignment led to enhanced Z-spectrum fidelity and increased metabolite quantification accuracy.

Conclusions:

• RECENet offers a robust and generalizable solution for contrast-invariant deformable registration of CEST MRI.
• The framework effectively addresses challenges posed by inter-offset contrast variations.
• RECENet significantly improves both spatial alignment and quantitative accuracy in CEST MRI analysis.