In Magma (New York, N.Y.)
OBJECTIVE : To achieve efficient segmentation of human supraclavicular adipose tissue (sclavAT) using high-resolution T2-weighted magnetic resonance images.
METHODS : High-resolution 1.0 mm isotropic 3D T2-weighted images covering human supraclavicular area were acquired in transverse or coronary plane from 29 volunteers using a 3.0 T MRI scanner. There were typically 144/288 slices for the transverse/coronary scans for each subject, which amounts to a total of 6816 images in 29 volunteers. A U-NET network was trained to segment the supraclavicular adipose tissue (sclavAT). The performance of the automatic segmentation method was evaluated by comparing the output results with the manual labels using the quantitative indices of dice similarity coefficient (DSC), precision rate (PR), and recall rate (RR). The auto-segmented images were used to calculate the sclavAT volumes and registered to the MR fat fraction (FF) images to quantify the fat component of the sclavAT area. The relationship between body mass index (BMI), the volume and FF of sclavAT area was evaluated for all subjects.
RESULTS : The DSC, PR and RR of the automatic sclavAT segmentation method on the testing datasets were 0.920 ± 0.048, 0.915 ± 0.070 and 0.930 ± 0.058. The volume and the mean FF of sclavAT were both found to be strongly correlated to BMI, with the correlation coefficient of 0.703 and 0.625 (p < 0.05), respectively. The averaged computation time of the automatic segmentation method was approximately 0.06 s per slice, compared to more than 5 min for manual labeling.
CONCLUSION : The present study demonstrates that the proposed automatic segmentation method using U-Net network is able to identify human sclavAT efficiently and accurately.
Wu Bingxia, Cheng Chuanli, Qi Yulong, Zhou Hongyu, Peng Hao, Wan Qian, Liu Xin, Zheng Hairong, Zhang Huimao, Zou Chao
2022-Dec-20
Brown adipose tissue, Deep learning, Fat fraction, Magnetic resonance imaging, Supraclavicular adipose tissue
