An Interpretable Deep‐Learning Approach for Efficient CEST Parameter Quantification: Importance‐Ranked Saturation Transfer MRI Protocol

PURPOSE: An optimal design of saturation-transfer MR fingerprinting (ST-MRF) sequences is essential to accelerate imaging and improve tissue quantification accuracy. This study aims to develop an interpretable deep-learning framework, importance-ranking network (IRnet), which can rank and identify the most informative dynamic scans, enabling optimized acquisition with a minimal number of scans while maintaining quantitative accuracy. METHODS: IRnet was developed to learn the scan-specific contributions to the latent representation of tissue parameters derived from ST-MRF. It consists of an encoder network trained on ST-MRF signals and corresponding ground-truth tissue parameters simulated using three-pool Bloch-McConnell equations. A shallow network was then used to predict the latent tissue representations, enabling scan importance to be ranked based on the magnitude of the learned weights. RESULTS: IRnet achieved more than a two-fold reduction in acquisition time while maintaining good reconstruction accuracy, with a normalized root-mean-square error of 6.2% when compared to ST-MRF with a full range of dynamic scans as a reference. The method consistently outperformed the pseudo-random selection and the least absolute shrinkage and selection operator-based approach, particularly for challenging amide proton transfer (APT) parameters, proton exchange rates and pool size ratios, to which ST-MRF is less sensitive than to magnetization transfer contrast (MTC) and water parameters. The tissue parameters obtained from IRnet and reference sequences demonstrated excellent consistency. CONCLUSIONS: IRnet enabled efficient and accurate tissue quantification by selecting a sparse, informative subset of acquisition parameters. This interpretable data-driven approach achieved accelerated quantitative CEST imaging and holds potential for translation into clinical protocols.