Automatic deep learning-based segmentation of neonatal cerebral ventricles from 3D ultrasound images

Abstract

In comparison to two-dimensional (2D) ultrasound (US), three-dimensional (3D) US imaging is a more sensitive alternative for monitoring the size and shape of neonatal cerebral lateral ventricles. It can be used when following posthemorrhagic ventricular dilatation, after intraventricular hemorrhaging (IVH), which is bleeding inside the lateral ventricles of the brain in preterm infants. Tracking ventricular dilatation is important in neonates as it can cause increased intracranial pressure, leading to neurological damage. However, manually segmenting 3D US images is time-consuming and tedious due to poor image contrast and the complex shape of cerebral ventricles. In this paper, we describe an automated segmentation method based on the U-Net model for the segmentation of 3D US images that may contain one or both ventricle(s). We trained and tested two models, a 3D U-Net and slice-based 2D U-Net, on a total of 193 3D US images (105 one ventricle and 88 two ventricle images). To mitigate the class imbalance of the object vs. background, we augmented the images through rotation and translation. As a benchmark comparison, we also trained a U-Net++ model and compared the results with the original U-Net. When all the images were used in a single U-Net model, the 3D U-Net and 2D U-Net yielded a Dice similarity coefficient (DSC) of 0.67±0.16 and 0.76±0.09 respectively. When two 2D U-Net models were trained separately, they yielded a DSC of 0.82±0.09 and 0.74±0.07 for one ventricle and two ventricle images, respectively. Compared to the best previous fully automated method, the proposed 2D U-Net method reported a comparable DSC when using all images but an increased DSC of 0.05 when using only one ventricle image.