Liver segmentation using 3D CNNs with high level shape constraints

Automatic liver segmentation from abdominal computed tomography (CT) images is a fundamental task in computer-assisted liver surgery programs. Recently, deep convolutional neural networks (CNNs) are served as the first choice in many volumetric segmentation tasks. However, the most used cross entropy loss treats each pixel independently and equally, which makes the network sensitive to fuzzy boundaries and heterogeneous pathologies. To address these issues, we propose an automatic segmentation framework based on a 3D CNN with a hybrid loss function. The hybrid loss function consists of three parts. The first part is an adaptively weighted cross entropy loss, which pays more attention on misclassified pixels. The second part is an edge-preserved smoothness loss, which guarantees neighbouring pixels with the same label have similar outputs, while neighbouring pixels with different labels have dissimilar outputs. The third part of loss is a shape constraint used to model high level structure differences. In our experiments, data augmentation is performed both in the training stage and the test stage. Finally, a conditional random field model is used to refine the segmentation results. We extensively evaluated our method on two datasets: the Segmentation of the Liver Competition 2007 (SLIVER07), and the Combined (CT-MR) Healthy Abdominal Organ Segmentation (CHAOS) Challenge.