Automatic Fetal Fat Quantification from MRI

Normal fetal adipose tissue (AT) development is essential for perinatal well-being. AT, or simply fat, stores energy in the form of lipids. Malnourishment may result excessive depleted adiposity. Although previous studies showed a correlation between amount AT and outcome, prenatal assessment limited by lacking quantitative methods. Using magnetic resonance imaging (MRI), 3D fat- water-only images entire fetus can be obtained from two-point Dixon to enable lipid quantification. This paper first present methodology developing deep learning (DL) based method fat segmentation on MRI. It optimizes radiologists’ manual delineation time produce annotated training dataset. consists two steps: 1) model-based semi-automatic segmentations, reviewed corrected radiologist; 2) automatic using DL networks trained resulting Segmentation 51 fetuses was performed with method. Three were trained. We show significant improvement times (3:38 h $$\rightarrow \,{<}$$ 1 h) observer variability (Dice 0.738 $$ 0.906) compared segmentation. Automatic 24 test cases Residual U-Net, nn-UNet SWIN-UNetR transformer yields mean Dice score 0.863, 0.787 0.856, respectively. These results are better than variability, comparable adult pediatric A Radiologist six new independent segmented best performing network (3D U-Net), 0.961 significantly reduced correction 15:20 min. these novel methods short MRI acquisition time, whole body subcutaneous lipids quantified individual clinic large-cohort research.