Dictionary building is a key component of Magnetic Resonance Fingerprinting (MRF) and the step sizes of the parameters simulated typically determines the quanta of the parameters being determined. Current work involves demonstration of Partial Least Squares (PLS) regression as a general framework for rapid and noise tolerant dictionary matching for MRF¹. PLS consists of Principle Component Analysis (PCA) and regression analysis^{2, 3.} PLS provides the intermediate values not defined as part of the dictionary steps as it involves a linear summation of regression coefficients. In addition, it also performs denoising (PCA part of PLS) of the maps in the presence of low SNR relevant to acquisitions like MRF. PLS provides the regression coefficient matrix and hence iterative search is not necessary. This provides for fast dictionary matching. PSIF is a Steady State Free Precession (SSFP) based sequence. The echo intensity of PSIF sequence can be obtained using the analytical equation using equation (17) from ref. ⁴. The dictionary was generated using the analytical equation with TR/alpha ranging from 0-100ms/0-900 and T₁ and T₂ ranging from 100-5000ms and 20-2000ms. The steps size selected for dictionary matching for T₁ was 100 to 2000 in steps of 20 , 2000 to 5000 in steps of 300. The steps size selected for T₂ was 20 to 100 in steps of 10, 100 to 200 in steps of 50 and 200 to 3000 in steps of 200. Three data sets with 109 brain images were acquired from Siemens Avanto 1.5T scanner with TR/alpha ranging from 15-66ms/170-350 respectively with minimum echo time (TE). TR/alpha combinations were considered as predictor and T₁/T₂ were the response variables. The response variables were computed based on 109 TR/alpha combinations. Different combinations of T₁/T₂ based on the step size employed to build the dictionary. Regression co-efficient matrix (β) was obtained through training on this dictionary. Each signal evolution obtained from the scanner was multiplied with the regression co-efficient matrix for localizing different brain tissues. PLS was tested for its ability of matching by considering different Signal-to-Noise Ratio (SNR) images as compared to traditional dot product matching for in vivo data. Gaussian noise was simulated for required SNR values of 8, 11 and 15db as in ref 5. The noise generated for required SNR was introduced at different time intervals in the data. Noise was introduced to the data at three different intervals i.e., the first 10 images, time points 51 - 60 and last 10 images. The noise-corrupted data was then used for map generation using both PLS and dot product methods.The T₁ and T₂ maps obtained through PnP-DP [6] matching is as shown in the first column of figure 1. The second column shows the T₁ and T₂ maps obtained through PnP-PLS method. It can also be noticed that the relaxation time values lie in the physiological ranges for these tissues. The values of T₁ &T₂ obtained for different tissue types of the brain through PnP-PLS is as shown in the table 1. It can be observed that these values are similar to previously published values for these parameters. The results for Gaussian noise simulated for PnP-PLS and PnP-DP is as shown in figure 2. The qualitative results show that at lower SNR levels PnP-PLS can match better compared to PnP-DP. It can also be observed that PnP-PLS can match better in the presence of noise at the intermediate and end of the signal evolution as compared to the beginning of the signal evolution. This might be attributed to the relatively low TR combinations employed in PnP-DP during the first 10 acquisitions and hence inherently low SNR data. Table 2 and table 3 shows the values T₁& T₂of different brain tissue types at various noise levels Higher contrast between grey and white matter in PnP-PLS can be attributed to the intermediate values provided by regression coefficients through PLS matching. This enables determination of parametric values lying between the discrete steps of the dictionary. PCA component of PLS has the ability to denoise the data and perform linear regression significantly better than the straight forward dot product. Performance of PLS on retrospective noisy data is validated by the results shown in figure 2. Current and future work involves application of this method to derive partial volume of gray matter, white matter and CSF within a given voxel and generate probability maps of these quantities.