Magnetic Resonance Fingerprinting (MRF) framework can be used in acquiring quantitative mappings of multiple parameters (i.e. T₁, T₂ and proton density) in a very rapid way [1]. The huge data provided by MRF contains abundant biological information including tissue composition. The signal evolution in a single voxel can be regarded as a superposition of the signal from several tissue types, such as gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF). Based on modeling this multicomponent nature of signal, the relative fractions of each tissue can be obtained, which is of potential diagnosis value [2]. In order to decompose signal, a multi-component model should be built and it requires a priori knowledge of T₁ and T₂ of each component. However, in practice the system instability and individual differences can make the selection of T₁ and T₂ problematic, and inaccurate parameters can lead to biased fractions as well as great fitting residuals. To optimize the selection of T₁ and T₂ for decomposing brain tissue, we propose to correlate the signals with themselves in a data-driven way and cluster them based on the correlation map. The T₁ and T₂ are extracted from the centers of each cluster as the optimized value. Using both phantom and in-vivo measurement, we demonstrate that the tissue fraction maps based on the proposed method can provide both good tissue differentiation and improved robustness.MRF measurements were performed on a 3T system (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany) with an inversion-prepared FISP MRF sequence [3]. The TR varied from 10 to 12 ms and the flip angle varied from 5 to 80 degrees. The total scanning time was 10s. A PVP water solution phantom was made with concentration of 0%, 5%, 10%, 15%, 20%, 30% of PVP in separate compartments respectively. The cross-correlation are performed for each pair of voxels. The calculated correlation coefficients are used as the distance function for fuzzy c-means (FCM) method. In every step of FCM, the centers of clusters are adjusted until they can represent the members optimally. When the clustering is completed, each center is matched to MRF dictionary so that their T₁ and T₂ can be obtained. We first cluster the voxels into five categories and then choose three of them that can best represent WM, GM and CSF. The extracted T1 and T2 are then used to decompose MRF signal with a three-component model as described in [1]. Our methods are compared with the method proposed in the original MRF paper [1], in which the signal model uses empirical values of T₁/T₂ for each type of tissue: white matter (660/70 ms), gray matter (1200/90 ms) and CSF (5000/700 ms). The relative residual is calculated as 2-norm of fitting residual divided by the 2-norm of signal.The phantom result shows that the similarity between signal are negatively correlated with the concentration difference, as clearly demonstrated in the correlation map (Fig. 1b). The diagonal positions provide the intra-sample similarity and are much brighter than the rest that represent the inter-sample similarity. The darkest blocks are at the left bottom and right top corners, which represents the dissimilarity between 30% PVP and water. The brain tissues are clustered according to their correlation into five types (Fig. 2). From their T₁ and T₂ values and spatial distribution, they can be identified as pure WM (Fig. 2c), pure GM (Fig. 2d), pure CSF (Fig. 2g) and two partial volume types of GM and CSF (Fig. 2e and 2f) respectively. The T₁/T₂ of pure WM (800/90 ms), GM (1100/110 ms) and CSF (3500/2000 ms) are chosen as the optimized value to do decomposition. Compared to the anatomic picture (Fig. 3g), both our and Ma D et al.’s method can provide a good contrast between WM and GM as shown in Fig. 3a and 3d. In the forehead, some parts of gray matter are missing in the previous method but not in our image (Fig. 3b and 3e). The relative residual of our method is also lower, which means less anatomic information is lost (Fig. 3h and 3i). We believe these improvements come from the use of optimized T₁ and T₂ to represent the signal rather than empirical values. We also tried more than five clusters in the first step but there is no further improvement.Compared to the original method of using a model of empirical T₁ and T₂, the proposed method makes full use of the similarity between signal in voxels, which results in a more robust brain segmentation and tissue fraction estimation.