Completeness as Correctness
As well as conventional image understanding including partial understanding of medical images, the analysis results from complete understanding of the scheme must be correct. That is, the anatomical label(s) must be assigned to every voxel without errors. However, the correctness is strongly dependent on the classifications used for anatomical labeling.
An important characteristic in classification of anatomical objects is that there are hierarchical structures of the classes. For instance, the liver can be recognized on several levels such as the entire liver, segments defined by Couinaud [129, 130], tissues, and cells. Such levels of detail required in medical image understanding may depend on the purposes, which are also restricted by spatial resolution of images.
Basically, the stance of complete medical image understanding is to make the maximum effort to classify in as much detail as possible. However, considering the computing time cost in clinical situations, a practical level as far as matching the needs can be chosen. As a matter of fact, it should be achieved by the balance of maximum detail of the classification and the correctness of the results. For example, if we pursue only the correctness, the classification can be the two classes of “human body” and “air (rest of the human body).” In most cases, this classification is nonsense for any diagnostic or therapeutic purposes. Instead, cell-level classification in CT images with a spatial resolution of submillimeter level is not feasible. In other words, the classification completeness should be defined adaptively depending on several factors, such as imaging modality, image resolution, purpose, and computational cost permissible to the purpose. This is an important discussion for practical and application-oriented aspects of computational anatomy.
The remainder of this book details complete medical image understanding, including how to achieve the results.