Joint Classification-Regression Forests for Spatially Structured Multi-Object Segmentation

In many segmentation scenarios, labeled images contain rich structural information about spatial arrangement and shapes of the objects. Integrating this rich information into supervised learning techniques is promising as it generates models which go beyond learning class association, only. This paper proposes a new supervised forest model for joint classification-regression which exploits both class and structural information. Training our model is achieved by optimizing a joint objective function of pixel classification and shape regression. Shapes are represented implicitly via signed distance maps obtained directly from ground truth label maps. Thus, we can associate each image point not only with its class label, but also with its distances to object boundaries, and this at no additional cost regarding annotations. The regression component acts as spatial regularization learned from data and yields a predictor with both class and spatial consistency. In the challenging context of simultaneous multi-organ segmentation, we demonstrate the potential of our approach through experimental validation on a large dataset of 80 three-dimensional CT scans.