Automatic Segmentation of the Cerebral Ventricle in Neonates Using Deep Learning with 3D Reconstructed Freehand Ultrasound Imaging

Preterm neonates can be subject to ventricu-lomegaly, which is an enlargement of the cerebral ventricle system (CVS) that can lead to brain damage. In clinical practice, 2D coronal hand-held ultrasonographic scans are performed to assess CVS dilation. Estimating CVS volumes from 2D images is, however, imprecise and time consuming since 3D information is lacking. To address this issue, we propose a 3D reconstruction method and an automatic deep learning segmentation algorithm. The accuracy of the 3D reconstruction was assessed by calculating Mean Absolute Distance (MAD) between manual segmentation of the corpus callosum (CC) on a ground reference and the 3D reconstructed volume, a mean value of 1.55 mm was obtained. The accuracy of the segmentation was evaluated using Dice, Hausdorff distance (dH) and MAD, respective average values of 0.816, 13.6 mm and 0.62 mm were obtained. The computation time of a segmentation for one 256 × 256 × 256 volume was 5 s.