Prior based Cascading of Attention Unets for Segmentation to Estimate Fat and Lean Mass from 3D Fetal Ultrasound Imaging

The ultrasonography during pregnancy is performed predominantly to assess various aspects of fetal growth. A reliable estimate of fat vs lean mass ratio in fetus can reveal subtleties of fetal development and possible abnormalities. The fat and lean mass can be quantified by analysing 2D limb cross-sections extracted from 3D ultrasound scans. Presently, this assessment is performed manually and also requires meticulous effort. This work proposes a novel deep learning based approach to automatically segment the regions of fat and lean mass from limb volume slices. The proposed method uses two attention U-Nets in cascaded mode, the output of one being used as prior for another. The technique achieves average validation dice coefficient as 0.8533 on dataset comprising slices from 39 limb volumes. The results indicate significant performance gain of this system as compared to a single attention U-Net.