Ensemble Learning with CA
Ensemble learning, such as AdaBoost  and random forest , has been prominent in recent segmentation research. Sophisticated segmentation examples by ensemble learning have been presented so far. A possible limitation is, however, that the segmentation result tends to be unnatural in terms of shape because the segmentation is performed in dependently voxel-by-voxel. A number of methods have been proposed to make the segmentation results natural in shape. The combination of MRF with AdaBoost , Spatial Boost , and Spatial AdaBoost  are typical examples. In [176, 177], a wide range of spatial information was employed for segmentation of subcortical structures in brain images and shape-based retrieval in cluttered images. However, in the context of medical image analysis, it is important to deal with organ- or tissue-specific features in the boosting- based segmentation.
A boosting algorithm that can take into account a target object’s specific shape prior was presented in , in which a new shape loss function evaluating directions of normal vectors of a target surface was proposed and minimized together with a conventional error loss function in the training process. The trained ensemble classifier was applied to extract lung fissures that are the thin pleural- covered potential spaces separating the lung lobes. Figure 2.27 shows an original CT image and the difference in outputs before binarization between a classifier trained by AdaBoost and that by Shapeboost. It is found from the figure that the shape of the algorithmically more enhanced area by Shapeboost is more similar to true fissures than those using Adaboost. The segmentation results of graph-cuts based on the enhanced results is satisfactory as shown in part (e).
Another alternative approach incorporates a novel shape loss function that evaluates distance between an extracted shape and a subspace of SSM of an organ and minimizes a total loss function including not only a conventional error loss but also the proposed shape loss . The method was successfully applied to spleen
Fig. 2.27 Outputs of ensemble classifiers trained by boosting algorithms without and with the proposed shape loss. (a) axial section of an original CT volume, (b) algorithmically more enhanced lung fissures by the proposed Shapeboost and (c) by AdaBoost. (d) sagittal section of an original CT volume and (e) algorithmically enhanced fissures by the proposed Shapeboost overlaid with the lung lobe segmentation result using a graph-cuts algorithm with an energy term derived from the algorithmically enhanced fissures 
segmentation and validated using 80 CT volumes. Details on the topic of spleen segmentation, in which the ensemble learning is applied, can be found in Sect. 3.9.3.