Segmentation of Intraoperative 3D Ultrasound Images Using a Pyramidal Blur-Pooled 2D U-Net.

Automatic localization and segmentation of the tumor and resection cavity in intraoperative ultrasound images can assist in accurate navigation during image-guided surgery. In this study, we benchmark a pyramidal blur-pooled 2D U-Net as a baseline method to segment the tumor and resection cavity before, during, and after resection in 3D intraoperative ultrasound images. Slicing the 3D image along three transverse, sagittal, and coronal axes, we train a different model corresponding to each axis and average three predicted masks to obtain the final prediction. It is demonstrated that the averaged mask consistently achieves a Dice score greater than or equal to each individual mask predicted by only one model along one axis.