Anatomical objects have been classified into vast classes of objects from organ level to cell level (Fig. 1.15). However, when we introduce the notion of classes to medical image understanding, the feasibility of classification strongly depends on the spatial resolution of a given image. To achieve true completeness in classification by covering all classes, almost infinite image resolution might be required in each image acquisition. This is almost impossible in clinical situations, and therefore, we need a practical solution, which is discussed later.
This is quite simple. In terms of the minimum unit of medical images, voxel, there must be no voxel unclassified. In other words, all the voxels in a given image must be anatomically labeled (exhaustive labeling) according to the anatomical classification including surrounding air and even artificial objects such as cannulas or catheters. Considering the partial volume effect, a single voxel can be a mixture of several anatomical structures, and therefore, multiple labels may need to be assigned to a single voxel.