Joint Supervoxel Classification Forest for Weakly-Supervised Organ Segmentation

This article presents an efficient method for weakly-supervised organ segmentation. It consists in over-segmenting the images into objectlike supervoxels. A single joint forest classifier is then trained on all the images, where (a) the supervoxel indices are used as labels for the voxels, (b) a joint node optimisation is done using training samples from all the images, and (c) in each leaf node, a distinct posterior distribution is stored per image. The result is a forest with a shared structure that efficiently encodes all the images in the dataset. The forest can be applied once on a given source image to obtain supervoxel label predictions for its voxels from all the other target images in the dataset by simply looking up the target’s distribution in the leaf nodes. The output is then regularised using majority voting within the boundaries of the source’s supervoxels. This yields sparse correspondences on an over-segmentation-based level in an unsupervised, efficient, and robust manner. Weak annotations can then be propagated to other images, extending the labelled set and allowing an organ label classification forest to be trained. We demonstrate the effectiveness of our approach on a dataset of 150 abdominal CT images where, starting from a small set of 10 images with scribbles, we perform weakly-supervised image segmentation of the kidneys, liver and spleen. Promising results are obtained.