Accelerated Chemical Exchange Saturation Transfer Imaging With Deep Unrolling Networks and Synthetic Brain Tumor Datasets

To develop a model-based deep unrolling network for high-quality image reconstruction of accelerated multi-channel chemical exchange saturation transfer (CEST) imaging.

Inspired by the theory of model-based deep learning (MoDL), we unrolled the alternating direction method of multipliers (ADMM) optimization for image reconstruction into a network, named MoDL-ADMM. Additionally, we designed a CEST image synthesis pipeline (BraTS-CEST) to obtain large-scale brain tumor training data using open BraTS and fastMRI datasets and Bloch-McConnell simulations. The performance of the proposed MoDL-ADMM method was evaluated on data from healthy volunteers and brain tumor patients using retrospective and prospective undersampling with various acceleration rates. We compared the reconstruction results of MoDL-ADMM with the original MoDL and other methods, including the state-of-the-art CEST-VN.

The proposed BraTS-CEST dataset yielded high-quality CEST images compared to previous methods and reduced the reconstruction error of the trained networks. As the acceleration rates increased from 3 to 6, MoDL-ADMM consistently reconstructed accurate source images and amide proton transfer-weighted (APTw) maps, outperforming GRAPPA, L + S, the original MoDL, and CEST-VN. The ablation studies further validated the effectiveness of the structural design, particularly the selective kernel networks and the learnable sparse transformation.

The proposed MoDL-ADMM, trained with the BraTS-CEST synthetic dataset, effectively reconstructed high-quality CEST source images and APTw maps from undersampled multi-channel data.