MR elastography (MRE) is used to estimate hepatic tissue stiffness in vivo. A continuous harmonic vibration is applied to the upper abdomen, and shear waves are generated that propagate through the liver. The resulting wave images are analyzed with an inversion algorithm to generate an elastogram that depicts hepatic stiffness spatial distribution¹. Mean values within regions of interest (ROIs) placed on the elastogram representing mean liver stiffness may be calculated. These mean hepatic stiffness values have been shown to accurately detect advanced fibrosis in a variety of liver conditions, but they are only moderately accurate in detecting earlier-stage fibrosis^{2, 3}. Through our lab's use of an advanced 3D version of MRE (3D MRE) in pediatric research studies, we have noticed qualitatively that the distribution of hepatic stiffness is homogeneous in some children, but heterogeneous in others with patchy areas of elevated stiffness against a background of lower stiffness. Although the cause of this apparent stiffness heterogeneity is unknown, it plausibly may represent true heterogeneity in underlying liver fibrosis that might not be well characterized by mean hepatic stiffness. We hypothesized that local heterogeneity in the distribution of hepatic stiffness may be helpful in detecting fibrosis. Thus, the purpose of this study was to assess the feasibility of using heterogeneity information in the distribution of hepatic stiffness to improve the performance of 3D MRE in the detection of liver fibrosis, especially earlier-stage fibrosis.This retrospective analysis of data acquired prospectively between 2011 and 2014 included 70 children (53 boys, 17 girls, mean age 13 yrs, range 10-18 yrs, mean body mass index 30 kg/m², range 21-51 kg/m²) with known or suspected NAFLD. Parents gave informed consent. Children gave informed assent and underwent 3D MRE at 60Hz within 180 days of clinical-care liver biopsy. MRE scans were performed on a 3T GE scanner using a motion-sensitized spin-echo echo-planar imaging sequence⁴. Biopsies were scored for fibrosis on a 5-point nominal scale (0-4) according to NASH CRN criteria⁵. Subjects were dichotomized into two classes: 'no fibrosis' (stage 0) vs. 'any fibrosis' (stage 1-4). Two elastogram-derived metrics were used to discriminate 'no fibrosis' from 'any fibrosis': mean hepatic stiffness, and a support vector machine (SVM)-generated score designed to account for local hepatic heterogeneity. Mean hepatic stiffness values were calculated from ROIs placed on elastograms. For the SVM-generated score, a bi-Gaussian model was fitted to the histogram of hepatic stiffness values, and the following features of the histogram of hepatic stiffness values were recorded for each of the two Gaussian models: mean stiffness values, standard deviation of stiffness values, and number of voxels. An SVM model with a radial basis kernel function was developed based on the features derived from the bi-Gaussian model and the known class label. The parameters for the SVM model were determined based on a least average classification error from 10-fold cross-validation procedures. For each child’s elastogram, the final SVM model computed a score ranging from 0 to 1, with 1 being highly likely that the child has fibrosis. Performance of classification was determined for each of the two elastogram-derived metrics (i.e. mean hepatic stiffness, and SVM-generated score) using the receiver operating characteristic (ROC) curve analyses. The measure of overall performance was the area under the ROC curve (AUROC).The distribution of fibrosis stages 0, 1, 2, 3, and 4 for the 70 enrolled children were 44 (63%), 17 (24%), 4 (6%), 4 (6%) and 1 (1%), respectively. Informally comparing AUROC for the two metrics, the SVM-generated score had better overall classification performance than mean hepatic stiffness (0.87 vs. 0.78). Comparing the ROC curve for the two elastogram-derived metrics, the SVM-generated score had higher sensitivity for detecting liver fibrosis for any given specificity greater than 0.75.We hypothesize that the SVM-generated score has higher sensitivity because incorporated within that score is information from the bi-Gaussian model which characterizes heterogeneity in the distribution of hepatic stiffness. For children with fibrosis, more than 85% were stage 1 or 2, suggesting that the SVM model helped improve the detection of earlier-stage fibrosis in this pediatric cohort. We demonstrated the feasibility of utilizing the heterogeneity in the distribution of hepatic stiffness in 3D MRE to improve detection of liver fibrosis, especially earlier-stage fibrosis.