Deep Cortical Vessel Segmentation Driven By Data Augmentation With Neural Image Analogy

During a craniotomy, a bone flap is temporarily removed from the skull to reveal the brain for surgery. The cortical vessels located at the surface of the brain are considered strong features to guide surgeons during the procedure, since they appear in both preoperative and intraoperative images and are an indication of how the brain may have shifted. We propose a method utilizing a deep neural network to perform cortical vessel segmentation in craniotomy images captured through the surgical microscope. Following a U-Net architecture, our solution classifies each pixel of a craniotomy image as vessel, parenchyma, or surrounding tissue and background. We use neural image analogy to build a diverse training set of unique images mirroring cortical anatomy generated from a limited amount of manually labeled data. The synthesized images enhance generalization of our model to various types of cortical surface appearances and geometries. Experiments on real data from human patients show that intraoperative cortical vessel segmentation can be performed accurately.