Brown and white adipose tissues play opposite functions in whole body metabolism by burning and storing energy. The recent identification of Brown adipose tissue (BAT) in adults [1] and its ability to normalize metabolic disorders has triggered a large interest in implementation of methods for characterizing and quantification. As interscapular region has a heterogeneous mixture of BAT and WAT, there is a need for increasing the specificity of these compartments by imaging techniques that will be suitable for longitudinal studies. In this study we have implemented a multiparametric MR imaging and evaluated different segmentation algorithms like multidimensional thresholding, region growing, clustering, and machine learning approach for its suitability and efficacy to separate WAT from BAT depots based on MR features fat-fraction (FF), T₂ and T₂*.

In vivo imaging and image processing: Twelve C57BL/6 (23 ± 1.8 g) mice and 8 male Wistar rats of 7 and 11 weeks were subjected to imaging on Varian 9.4T/31 cm and Bruker Clinscan MRI systems respectively. Dixon imaging was utilized for computing fat fraction and relaxation maps T₂ and T₂* were obtained by multi-gradient echo and multi-spin echoes respectively. The data processing is as shown in Fig.1 which included a) MR imaging based calculation of fat-fraction (FF), T₂ and T₂* maps, b) data analysis and characterization of different tissues and c) image segmentation and result analysis. Four segmentation methods [2]: multidimensional thresholding (MTh); region-growing (RG); fuzzy c–means (FCM) and neural-network (NNet) were evaluated on the interscapular region and validated against manually defined BAT, WAT and muscle. FF, T₂ and T₂* values formed the input feature vector for segmentation algorithms. A two-layer feed-forward network with 10 hidden neurons and sigmoid transfer function, and an output layer with softmax transfer function were used of training and testing. Scaled conjugate gradient based learning/training, and cross-entropy based performance evaluation was adopted in the study. Three different sized multi-parametric training sets (large, medium and small) were used to train the network. The training phase also included the validation and testing with 30% of the training samples. Mean squared error between the outputs and target vectors was used to measure the performance of the system. The input feature vector comprised of FF, T₂, and T₂* values and the output classes were BAT, WAT, muscle and background for the MR. Algorithms were developed in-house for i) background removal, ii) statistical & spatial method based outlier detection [3], iii) data characterization and iv) image segmentation using MATLAB, JMP and Medcalc software.

Statistical analysis of BAT segmentation on mice yielded a median Dice-Statistical-Index, and sensitivity of 89%, 92% for NNet, 82%, 86% for FCM, 72%, 74% for RG and 72%, 70%, for MTh, respectively and based on these results only NNet was used to segment the BAT in rats which had 97.1%, 96.6% and 99% median DSI, for BAT, WAT and muscle respectively. Decrease in FF values was observed in cold acclimatized animals (Fig. 2). NNet based segmentation in a representative slice is shown in Figure 3.

We have demonstrated that NNet based segmentation on multi-parametric MR images (FF, T₂, T₂*) achieves more specificity compared to other algorithms. It is also a reliable technique to differentiate BAT from surrounding tissues in the interscapular region of rats and mice. This method could facilitate characterization, quantification and longitudinal measurement of BAT in preclinical-models and can be potentially extended for human MR images.