Introduction

As an important endocrine organ, adipose tissue modulated by cerebral nervous system secrets various of hormone, which interacts with other endocrine organs and regulates body-weight¹. Excess storage of fat disrupts metabolic balance, leading to obesity and obesity-related metabolic syndrome. As a result, accurate quantification and segmentation of whole-body adipose tissue is critical for obesity management. In comparison with traditional qualitative anthropometric indices, including body mass index (BMI) and waist-to-hip-ratio, direct body composition quantification using magnetic resonance imaging (MRI) is more objective.
Several MR methods with manual, semi-automatic or automatic segmentation algorithms have been used to assess body fat deposition by exploiting the differences in the longitudinal relaxation time (T1) or chemical shift between adipose and aqueous tissues^2-3. However, these methods have only focused on part of the body, i.e., they have not been extended to the entire body.
In this study, a fully automatic segmentation method is proposed using the machine learning network model based on the proton density FF (PDFF) images for whole-body fat segmentation. The U-Net model is designed to differentiate SAT and IAT in the PDFF images. The proposed model shows excellent segmentation performance in several volunteers using a relatively small amount of training data.

Methods

All data were collected using 6-echo GRE sequence on a 3.0T clinical MRI scanner (uMR790, Shanghai United Imaging Healthcare, Shanghai, China). Twenty subjects were recruited with informed consent under institutional review board approval to participate in the whole-body MR scans. Each whole-body scan was performed in transverse slices covering the neck to knee areas with eight or nine beds based on the height of the subjects.
The six-echo GRE images were processed in MATLAB (NATICK, CA, USA) using a fat-water separation algorithm previously proposed by our research group to produce PDFF⁴. Then, the TAT was identified by thresholding the whole image with PDFF > 80%. Inside the TAT area, the SAT areas were labeled manually by ITK-SNAP slice by slice in all datasets. The remaining adipose areas were treated as the IAT. The flowchart of the image processing and data preparation procedures is shown in Fig. 1. A total of 906 PDFF images were obtained from all subjects, and these 906 images were randomly divided into training (504 images; 55.63%), validation (168 images; 18.54%), and test (234 images; 25.83%) sets.
The U-Net architecture is shown in Fig. 2. The entire network comprises twenty-three convolutional layers (including eighteen 3 × 3 convolution layers, four 2 × 2 up-convolution layers, and one 1 × 1 convolution layer) and four 2 × 2 maximum pooling operations. The same padding is used in the convolution operation. Copy processing (Fig. 2) is typically used to fully recover the refined spatial information. In this study, three indices were used to evaluate segmentation performance, i.e., the dice similarity coefficient (DSC), precision rate (PR), and recall rate (RR).

Results

The training time was approximately four hours for the training set (i.e., 504 images), and segmentation time was less than 8 s per 100 images (<0.08s / image). Table 1 shows the mean and standard deviation of DSC, PR, and RR for the validation and test sets with values of 0.976 ± 0.048, 0.978 ± 0.048, and 0.978 ± 0.050, respectively, in both validation and test sets.
The segmentation results for two subjects (one female, BMI: 19.05 kg/m2；one male, BMI: 26.03 kg/m2) are shown in Fig.3, where six transverse slices from neck to knee are shown for both subjects. The SAT images output from the network were compared to the manual labeled images, and only very slight differences were observed. Then, the IAT images were segmented by removing the SAT area output from the network model in the TAT images (Fig. 4).

Discussion and conclusions

In this paper, we have proposed an automatic segmentation method to discriminate SAT and IAT based on MR PDFF images. The proposed method represents an easy-use tool for whole-body fat distribution analysis.
In contrast to the most popular whole-body fat distribution analysis methods based on traditional T1 weighted spin echo images, which are susceptible to B1 inhomogeneities, the use of quantitative PDFF images avoids several system-related confounding factors, making the whole-body fat segmentation simple and robust. In comparison with previous methods, such as contour-based method⁵, which requires the initial snake contour be close to the real contour, and tedious iterations to find the final contour. However, the utilization of neural network makes the whole-body segmentation process fast and direct, and does not need any more manual inputs.
In our next work, we will explore the methods to classify the IAT into finer categories, e.g. brown adipose tissue, bone marrow, intra-/inter-muscular adipose tissue, etc.

Acknowledgements

This research was supported by the Natural Science Foundation of China (No. 61901462), the Guangdong Grant ‘Key Technologies for Treatment of Brain Disorders’ (2018B030332001) and Shenzhen Double Chain Grant ([2018]256)