Using Anatomical Priors for Deep 3D One-shot Segmentation

Abstract

With the success of deep convolutional neural networks for semantic segmentation in the medical imaging domain, there is a high demand for labeled training data, that is often not available or expensive to acquire. Training with little data usually leads to overfitting, which prohibits the model to generalize to unseen problems. However, in the medical imaging setting, image perspectives and anatomical topology do not vary as much as in natural images, as the patient is often instructed to hold a specific posture to follow a standardized protocol. In this work we therefore investigate the one-shot segmentation capabilities of a standard 3D U-Net architecture in such setting and propose incorporating anatomical priors to increase the segmentation performance. We evaluate our proposed method on the example of liver segmentation in abdominal CT volumes.