Ultrasound Segmentation Using a 2D UNet with Bayesian Volumetric Support

We present a novel 2D segmentation neural network design for the segmentation of tumour tissue in intraoperative ultrasound (iUS). Due to issues with brain shift and tissue deformation, pre-operative imaging for tumour resection has limited reliability within the operating room (OR). iUS serves as a tool for improving tumour localisation and boundary delineation. Our proposed method takes inspiration from Bayesian networks. Rather than using a conventional 3D UNet, we develop a technique which samples from the volume around the query slice, and perform multiple segmentation's which provides volumetric support to improve the accuracy of the segmentation of the query slice. Our results show that our proposed architecture achieves an 0.04 increase in the validation dice score compared to the benchmark network.