Automatic Segmentation of Antenatal 3-D Ultrasound Images

• Published in: IEEE transactions on biomedical engineering 2013-05, Vol.60 (5), p.1388-1400
• Main Authors: Anquez, Jeremie, Angelini, Elsa D., Grange, Gilles, Bloch, Isabelle
• Format: Article
• Language: English
• Subjects: 3-D ultrasonic imaging; Antenatal imaging; Bayes Theorem; biomedical image processing; Computational modeling; Databases, Factual; Female; Fetus; Histograms; Humans; Image segmentation; Imaging, Three-Dimensional - methods; level sets; Manuals; Models, Biological; Pregnancy; Statistics, Nonparametric; Studies; Ultrasonic imaging; Ultrasonography, Prenatal - methods
• ISSN: 0018-9294; 1558-2531
• DOI: 10.1109/TBME.2012.2237400

The development of 3-D ultrasonic probes and 3-D ultrasound (3DUS) imaging offers new functionalities that call for specific image processing developments. In this paper, we propose an original method for the segmentation of the utero-fetal unit (UFU) from 3DUS volumes, acquired during the first trimester of gestation. UFU segmentation is required for a number of tasks, such as precise organ delineation, 3-D modeling, quantitative measurements, and evaluation of the clinical impact of 3-D imaging. The segmentation problem is formulated as the optimization of a partition of the image into two classes of tissues: the amniotic fluid and the fetal tissues. A Bayesian formulation of the partition problem integrates statistical models of the intensity distributions in each tissue class and regularity constraints on the contours. An energy functional is minimized using a level set implementation of a deformable model to identify the optimal partition. We propose to combine Rayleigh, Normal, Exponential, and Gamma distribution models to compute the region homogeneity constraints. We tested the segmentation method on a database of 19 antenatal 3DUS images. Promising results were obtained, showing the flexibility of the level set formulation and the interest of learning the most appropriate statistical models according to the idiosyncrasies of the data and the tissues. The segmentation method was shown to be robust to different types of initialization and to provide accurate results, with an average overlap measure of 0.89 when comparing with manual segmentations.