Recent osteoarthritis (OA) research has aimed to identify disease biomarkers earlier than the current clinical standard of radiographic joint space narrowing. MRI relaxometry studies explain that T1ρ and T2 signal relaxation times reflect biochemical composition and could potentially isolate early OA biomarkers in OA affected cartilage¹. Traditionally, relaxometry studies have averaged relaxation time values across regions of interests based on specific anatomical landmarks (ROI); however, voxel based relaxometry (VBR) has been proposed as another method that could improve upon MRI analysis in arthritic cartilage². Despite the advantages to VBR, statistical analysis on the voxel level could be susceptible to noise, and VBR map homogenization would be difficult for articular cartilage. In this study we propose to use VBR significance thresholding in conjunction with cluster-extent based thresholding³ to define data-driven regions of interest (ROIs) that include the most critical information in Statistical Parametric Maps (SPM).96 subjects from a single OA cohort were considered for this study. It included 58 OA-affected patients (age=58.4±9.1 years; BMI=25.3±3.5 kg/m2; KL =1.63) and 36 control subjects (age=51.5±9.25; BMI=23.8±3.35 kg/m2; KL = 0). All imaging was done using a 3T MRI scanner (GE Healthcare, Milwaukee, WI, USA) with an 8-channel phased array knee coil (Invivo Inc, Orlando, FL, USA). Sagittal 3D T1ρ imaging sequences were obtained with the following parameters: TR/TE=9ms/min full, FOV=14 cm, matrix=256x128, slice thickness=4 mm, spin-lock frequency=500 Hz, time of spin lock (TSL) = 0/2/4/8/12/20/40/80 ms. Sagittal 3D T2 preparation time: TE=1.8/3.67.3/14.5/29.1/43.6/58.2. All images were morphed to the space of a reference using a previously proposed technique², after which, voxel-by-voxel statistics could be conducted, producing SPMs. Correlation between relaxation times and compartmental WORMS⁴ score for characterizing morphological cartilage lesions was performed obtaining Pearson R correlation Maps for T1ρ and T2. The maps were thresholded to identify regions with significant correlations (p<0.01) obtaining clusters of adjacent voxels. Voxels were considered to be adjacent if they had a line in common. Significance levels for cluster size were obtained through the permutation method described by Holmes et al³. T1ρ and T2 relaxation time values were averaged within the clusters and correlated to non-cartilage relevant OA metrics such as Bone Marrow Edema WORMS Score⁴, KOOS Pain Scores⁵. The same correlations were obtained using analogous relaxation averages computed in traditional articular knee cartilage ROIs (LT, LF, MT, LT, TrF, P, Figure 1), and results were compared.Five significantly sized clusters survived the primary threshold of WORMS cartilage lesion (significant correlation, p < 0.01) and the secondary threshold of adjacent voxel cluster size p < 0.05. Significant clusters from the T1ρ SPM were in the anterior tibiofemoral (ant-TRF-cl), lateral tibial and femoral (LF-LT-cl), medial tibia and femur (MF-MT-cl), patella (P-cl), and posterior tibiofemoral (pos-TRF-cl) compartments (Figure 1a). The cluster sizes were 883, 763, 368, 313, and 79 voxels respectively, and the lower limit (p = 0.05) for cluster size was 68 voxels. In the T2 SPM clusters populated in the lateral tibia and femoral (LF-LT-cl), patella (P-cl), medial tibia and femoral (MF-MT-cl), anterior tibiofemoral (ant-TRF-cl), and posterior tibiofemoral (pos-TRF-cl) compartments (Figure 1b). The cluster sizes were 572, 513, 262, 137, and 89 voxels respectively, and the lower limit (p = 0.05) for cluster size was 67 voxels. Figure 2 shows correlations of the average T1ρ relaxation time within each cluster to WORMS measured local BME. T1ρ SPM clusters generally had higher and more significant correlations (mean cluster R = 0.342, mean p =0.09; mean ROI R = 0.188, mean p =0.203), and the trend held true for T2 SPM clusters (mean cluster R = 0.433, mean p = 2.8e-04; mean ROI R = 0.13, mean p = 0.35). The clusters also outperformed traditional ROIs in correlation with KOOS Pain for T1ρ (mean cluster R = -0.23, mean p =0.039; mean ROI R = -0.16, mean p =0.16) and T2 (mean cluster R = -0.29, mean p =0.017; mean ROI R = -0.12, mean p =0.295).The proposed method shows that data driven different ROI selection although is seen in the compartments expected, show differences compared to traditional ROIs. Data-driven clusters show better relaxometry-to-outcome correlation values and significance when compared to ROIs. T1ρ and T2 values within our proposed clusters are significantly more associated with WORMS BME and KOOs Pain. This suggests that the data driven clusters better identify changes in cartilage biochemical composition, focusing analysis on the most informative areas of relaxation maps. Further experiments and longitudinal studies are necessary to exploit the performance of those clusters in the disease monitoring and outcome prediction.