Fast global interactive volume segmentation with regional supervoxel descriptors

摘要

In this paper we propose a novel approach towards fast multi-class volume segmentation that exploits supervoxels in order to reduce complexity, time and memory requirements. Current methods for biomedical image segmentation typically require either complex mathematical models with slow convergence, or expensive-to-calculate image features, which makes them non-feasible for large volumes with many objects (tens to hundreds) of different classes, as is typical in modern medical and biological datasets. Recently, graphical models such as Markov Random Fields (MRF) or Conditional Random Fields (CRF) are having a huge impact in different computer vision areas (e.g. image parsing, object detection, object recognition) as they provide global regularization for multiclass problems over an energy minimization framework. These models have yet to find impact in biomedical imaging due to complexities in training and slow inference in 3D images due to the very large number of voxels. Here, we define an interactive segmentation approach over a supervoxel space by first defining novel, robust and fast regional descriptors for supervoxels. Then, a hierarchical segmentation approach is adopted by training Contextual Extremely Random Forests in a user-defined label hierarchy where the classification output of the previous layer is used as additional features to train a new classifier to refine more detailed label information. This hierarchical model yields final class likelihoods for supervoxels which are finally refined by a MRF model for 3D segmentation. Results demonstrate the effectiveness on a challenging cryo-soft X-ray tomography dataset by segmenting cell areas with only a few user scribbles as the input for our algorithm. Further results demonstrate the effectiveness of our method to fully extract different organelles from the cell volume with another few seconds of user interaction.