Deep Learning based Segmentation for Assessment of Fractional Limb Volume in 3D Fetal Ultrasound Imaging

Fractional limb volume is a fetal soft tissue parameter that can be estimated from the 3D ultrasonography during 2
^nd
and 3
^rd
trimester of the gestation period. This parameter has a significant value in assessment of fetal weight, growth and nutritional status. The calculation of fractional limb volume from 3D ultrasound scans involves segmenting limb 2D cross-section images. This work explores convolutional neural network based segmentation models for this task and compares their performance. The three models viz. attention Unet, a Unet style model with Xception convolution and feature pyramid network are evaluated based on their performance on dataset consisting of slices from 65 volumes. The results indicate that feature pyramid network marginally outperforms the other two models with average validation dice coefficient as 0.9006. The experiments also show that the performance of all models can be improved through basic post-processing of output mask.