Learning to segment key clinical anatomical structures in fetal neurosonography informed by a region-based descriptor

Abstract. We present a general framework for automatic segmentation of fetal brain structures in ultrasound images inspired by recent advances in machine learning. The approach is based on a region descriptor that characterizes the shape and local intensity context of different neurological structures without explicit models. To validate our framework, we present experiments to segment two fetal brain structures of clinical importance that have quite different ultrasonic appearances—the corpus callosum (CC) and the choroid plexus (CP). Results demonstrate that our approach achieves high region segmentation accuracy (dice coefficient: 0.81  %    ±  0.06 CC, 0.76  %    ±  0.08 CP) relative to human delineation, whereas the derived automated biometry measurement deviations are within human intra/interobserver variations. The use of our proposed method may help to standardize intracranial anatomy measurements for both the routine examination and the detection of congenital conditions in the future.