Scribble-based weakly supervised method for segmentation of neonatal cerebral ventricles from 3D ultrasound

Compared to conventional two-dimensional (2D) ultrasound, three-dimensional (3D) ultrasound (US) images are a more sensitive alternative for monitoring size and shape of neonatal cerebral lateral ventricles when diagnosing intraventricular hemorrhaging (IVH). The ventricles need to be segmented by an expert to estimate the ventricular volume which can be expensive and difficult to obtain. In this paper, we describe a scribble-based weakly supervised segmentation method that trains only on non-expert made scribbles. We trained and tested two models, a vanilla 3D U-Net benchmark and a weakly supervised learning for medical image segmentation (WSL4MIS) method, on a total of 56 3D US images. For all experiments, the WSL4MIS method had a higher mean DSC and lower standard deviation than that of the baseline 3D U-Net when both used scribble data for training.