Fetal cranial segmentation in 2D ultrasound images using shape properties of pixel clusters

Abstract

The detection of cranial dysmorphisms during pregnancy is achieved by assessing the cranial shape from 2D ultrasound images of the fetal head. As such, several algorithms have been presented to automate this task due to the fact that segmentation of the fetal cranium from ultrasound images is a central problem in obstetric care which is complicated by fuzzy boundaries and variability in fetal position and head shape. In this paper, we introduce a machine learning framework that employs a novel feature set which incorporates local statistics and shape information about pixel clusters (or superpixels) within an image, and evaluate the performance of the feature set in the task of segmenting the cranial pixels in an ultrasound image using a random forest classifier. Our experiments show that the features derived from the shapes of the pixel groupings outperform powerful features such as Haar features and achieved a 97.22% segmentation accuracy when applied to the task of fetal cranial segmentation in ultrasound images.