Hierarchical K-means clustering method for accelerated Lorentzian estimation (KALE) in chemical exchange saturation transfer-magnetic resonance imaging quantification.

Background: Quantification of in vivo chemical exchange saturation transfer (CEST) magnetic resonance signals is challenging due to contamination from coexisting effects, including the direct water effect and asymmetric magnetization transfer. Fitting-based analysis allows the calculation of multiple types of signals from the line shape of Z-spectra. However, the conventional voxelwise method has several drawbacks, including its long computation time and its susceptibility to image noise and Z-spectra oscillations, and it is difficult to determine the initial fitting parameters. Methods: Herein, we propose a K-means clustering method for accelerated Lorentzian estimation (KALE) in CEST quantification. Briefly, voxels in CEST images are clustered into K groups according to their Z-spectra characteristics. A 'groupwise' fitting process is then performed with preset initial values, yielding a set of fitted spectra and fitted parameters for each group. With the updated initial values, each group is further clustered into subgroups, and groupwise fitting is performed again. This hierarchical K-means clustering and parameter updating process continues until the pixel number or intensity error meets the termination criteria. Voxelwise fitting could be further conducted to improve the quantification images (termed voxel-K) by utilizing the previous groupwise KALE results as the initial values (termed group-K). Results: Incorporated with Lorentzian difference (LD) quantification, KALE was first optimized and evaluated on 5 healthy human brain datasets at 3 Tesla. Compared with traditional voxel-by-voxel LD quantification, the computation times of group-K and voxel-K were significantly reduced by ~85% and ~70%, respectively (P<0.001). Furthermore, the group-K images exhibited better denoising performance than traditional LD and voxel-K. KALE was further validated on six ischemic rat brains acquired at 7 Tesla, with both LD_group-K and LD_voxel-K displaying almost identical contrast maps with traditional voxelwise maps. When incorporated with the five-pool Lorentzian fitting (LF), KALE exhibited an improved contrast-to-noise ratio (CNR) for amplitude maps of each pool [P=0.003, 0.015, 0.047, and 0.047 for amide, nuclear Overhauser effect (NOE), magnetic transfer (MT) and guanidine amine, respectively] and improved fitting goodness (P=0.033). Conclusions: KALE quantification provides comparable or even superior contrast maps to traditional voxelwise fitting, with significantly reduced computation time. The 'smart' and hierarchical voxel-clustering and parameter updating process of KALE may facilitate more preclinical and clinical CEST applications.