CEST MRI Z-Spectra Reconstruction Using Transformer-based Deep Learning Models and Under Sampled Spectrum Signatures

Chemical exchange saturation transfer (CEST) contrast MRI provides unique molecular contrast by probing proton exchange between solutes and bulk water, but its clinical translation is limited by long acquisition times required to acquire finely sampled Z-spectra bringing a tradeoff between acquisition time and frequency offset responses. In this preliminary study, we investigated the performance of transformer-based deep learning models to reconstruct fully sampled CEST spectra. We first validated the approach using a lightweight model on simulated Z-spectra, followed by a scaled high-capacity transformer for human MRI data to handle the increased data scaling and spectral variability. We evaluate the model reconstruction quality on simulated Z-spectra and the human MRI data using root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and cosine similarity for regression analysis. We use structural similarity index measure (SSIM) and peak signal to noise ratio (PSNR) to evaluate image-level reconstruction quality. The transformer approach accurately recovered missing spectral information with an MAPE of 3.41% and RMSE of 0.0169. The reconstruction maintained high structural fidelity with a mean SSIM of 0.9658 and spectral fidelity with Cosine Similarity of 0.9951. These results validate that the scaled transformer architecture ensures robust generalization and accurately captures complex spectral dependencies in clinical data. These findings highlight the potential of deep learning-based reconstruction to accelerate CEST MRI acquisition supporting more efficient clinical protocols.