White Matter Segmentation from Cranial Ultrasound Images based on Convolutional Neural Network

White matter damage (WMD) is one of the most common consequences of preterm newborns, which may cause long-term neurodevelopmental deficits, such as cerebral palsy, abnormal audio-visual function, cognitive impairment, etc. Segmentation of white matter plays an important role in WMD detection and intervention. Manual segmentation of white matter is tedious and may cause inter- or intra-observer variability. In this work, ultrasound images from 148 premature infants were segmented using three convolutional neural networks, FCN, Unet and residual-structured fully convolutional network (res-FCN). Each preterm newborn collected three cross sections images from ultrasound. By comparison, the results showed that res-FCN had the most evaluation metrics with the best performance: Precision 78.94%, AD 26.54% on the lateral ventricle anterior horn plane; Recall 80.62%, Precision 77.09%, AO 63.00%, DSC 75.95% on the coronal lateral ventricle body plane; Recall 86.00%, AO 71.10% on the occipital lobe plane.