Model-Directed Detection of Ribs in Chest Radiographs

We describe a computational model for the ribcage anatomy seen in a posterior-anterior chest radiograph. This model contains an executive program which examines a declarative structure and can invoke parts of a procedural structure. The declarative structure is in the form of a relational graph and is primarily used to represent gross geometrical relations between the ribs in the ribcage. The procedural structure contains low-level programs that are associated with parts of the declarative structure and which can find detailed representations of the ribs in the image. The executive operates incrementally with a cycle of three distinct phases. First, it develops a plan for the "best" rib to find next; secondly, it tries to find that rib; and thirdly, it evaluates its results, incorporating any new information into the declarative structure. Thus, it finds the ribs that stand out the most clearly in the image data first, and then brings this knowledge to bear on the more obscure ribs. There are many advantages to this model. The ribs found are identified by name. The model is able to find the ribcage even if the ribs are partially obscured by disease processes or where they are faint, as in a high KVP film. The method of describing low-level programs allows experimentation with different kinds of algorithms without changing the high-level model structure, and the high-level declarative structure allows the identification and correction of some mistakes