Abstract
Automated surface segmentation is important and challenging in many medicalimage analysis applications. Recent deep learning based methods have beendeveloped for various object segmentation tasks. Most of them are aclassification based approach, e.g. U-net, which predicts the probability ofbeing target object or background for each voxel. One problem of those methodsis lacking of topology guarantee for segmented objects, and usually postprocessing is needed to infer the boundary surface of the object. In thispaper, a novel model based on 3-D convolutional neural network (CNN) andConditional Random Fields (CRFs) is proposed to tackle the surface segmentationproblem with end-to-end training. To the best of our knowledge, this is thefirst study to apply a 3-D neural network with a CRFs model for direct surfacesegmentation. Experiments carried out on NCI-ISBI 2013 MR prostate dataset andMedical Segmentation Decathlon Spleen dataset demonstrated very promisingsegmentation results.