Described herein is a technology for facilitating computer-aided detection and image understanding. In one implementation, an input set of training images of a target structure, such as an anatomical structure, is received. The input set of training images is spatially realigned to different landmarks to generate multiple bags of training images. At least one of the multiple bags comprises substantially all the training images in the input set, but realigned to a landmark. The multiple bags of training images may be used to train a spatial ensemble of detectors, which can be employed to generate an output result by automatically detecting a target structure in an input image.