Segmentation of neonates cerebral ventricles with 2D CNN in 3D US data: suitable training-set size and data augmentation strategies

For its clinical potential, segmentation of the preterm neonate's Cerebral Ventricular System (CVS) in 3D ultrasound (US) data using convolutional neural networks (CNN) is an emerging field. Nevertheless, gathering manually annotated data to efficiently train a CNN is difficult. In this paper, we address the question of how many training volumes and which kind of artificial data augmentation strategies would be suitable for this application. We explored this topic by training a U-net with different training-set size and by using several artificial data augmentation strategies. In our set-up, accurate segmentation results (Dice ≥ 0.8) were obtained with only 9 training volumes. The use of artificial data augmentation improved significantly (p <; 0.05) the accuracy obtained without data augmentation when performing horizontal flips (between the right and the left brain hemispheres). The other types of data augmentation that we tried did not significantly improve U-net accuracy.