3D statistical shape models for automatic segmentation of the fetal cerebellum in ultrasound images

The cerebellum is an important structure to determine fetal development because its volume has a high correlation with gestational age. Manual annotation of the cerebellum in 3D ultrasound images (to measure the cerebellar volume) requires highly trained experts to perform a time-consuming task. To assist in this task, we developed a totally automatic system for the 3D segmentation of the cerebellum in ultrasound images of the fetal brain, using a 3D Point Distribution Model (PDM) obtained from another statistical shape model based on a spherical harmonics (SPHARMs) representation, which provides a very efficient basis for the construction of statistical shape models of 3D organs with a spherical topology. Our PDM of the fetal cerebellum was automatically adjusted with the optimization of an objective function based on gray level voxel profiles, using a genetic algorithm. An automatic initialization and plane selection scheme was also developed, based on the detection of the cerebellum on each plane by a convolutional neural network (YOLO v2). Our results of the 3D segmentation of 18 ultrasound volumes of the fetal brain are: Dice coefficient of 0.83 +/- 0.10 and Hausdorff distance of 3.61 +/- 0.83 mm. The methods reported show potential to successfully assist the experts in the assessment of fetal growth in ultrasound volumes.