Automated Deep Learning Segmentation of Neonatal Cerebral Lateral Ventricles from Three-Dimensional Ultrasound Images

Compared to two-dimensional (2D) ultrasound (US), three-dimensional (3D) US is a more sensitive alternative that can provide quantitative information for monitoring neonatal cerebral lateral ventricles. Currently, manual segmentation of the cerebral ventricles is time-consuming due to poor image contrast and complex ventricle shape. To this end, I developed a fast and automated deep learning segmentation method to segment neonatal cerebral lateral ventricles from 3D US images. My proposed segmentation method is a 3D U-Net ensemble model comprised of three U-Net variants. The dataset had 190 3D US images, of which 87 contained both ventricles and 103 contained one ventricle. Using 5-fold cross-validation, the ensemble performed best with a Dice similarity coefficient (DSC) of 0.746±0.088, absolute volumetric difference (VD) of 2.8±2.5cm^3, and mean absolute surface distance (MAD) of 0.89±0.30mm on the two-ventricle images. For the one-ventricles images, the model reported DSC, VD, and MAD values of 0.820±0.102, 4.3±4.3cm^3, and 1.37±1.64mm, respectively.