Development of a Statistical-Based Deformable Shape Model for Segmentation and Recognition of Semi-Rigid Objects in Complex Backgrounds

(a)	(b)	(c)
Figure 1: Segmentation results of applying the new statistical shape model to x-ray CT images of laboratory mice:  (a) skull/ear canal, (b) lungs/lower-heart boundary, (c) spine/kidneys

Research Objective/Motivation:
The motivation for this research is the need for an algorithm that performs automatic segmentation and recognition of semi-rigid objects within complex backgrounds. A semi-rigid object is one that exhibits a controlled shape variability across multiple instances of that object. In a statistical sense, controlled variability implies that the intra-class variation of object shape is less than the inter-class shape variation. The algorithm is expected to handle cases in which the object’s boundary may be faint, obscured, or partially missing, so the segmentation algorithm should have the capacity to handle such a situation. The algorithm should also take full advantage of available a priori information regarding the appearance (e.g. shape- and intensity-characteristics) of the object. Finally, to ensure practical utility of the algorithm in an automated application, it must generate some measure of confidence associated with the segmentation result. In summary, we seek a segmentation/recognition algorithm capable of handling:

-semi-rigid objects
-complex backgrounds
-faint, obscured, or partially missing object boundaries
-available a priori information on object appearance
-a requirement for a segmentation confidence metric

To parallel this description with an application example, consider the field of medical imaging where anatomical structures may be somewhat predictable in terms of appearance (e.g. organ shape and location), but variations are always encountered from subject-to-subject and within the same subject over time. Also, there are areas of the anatomy (e.g. the abdominal cavity) where the appearance, location, and background of structures are quite unpredictable, making the segmentation task even more intractable. The difficulties created by these variations in the organ and its background are, at least in part, why semi-rigid organ segmentation is not as well-studied as rigid organ (or rigidly-enclosed organ, i.e., brain) segmentation. Also, in many cases of medical image analysis there is an abundance of a priori information available in the form of the patient’s own historical records as well as imagery from a potentially large population of other patients that could be used for algorithm training, for example. Finally, for an automated medical image analysis task to be used in a clinical setting, there must be a dependable confidence measure generated by the algorithm to provide an opportunity for a clinician to intervene in the event of questionable results.

Methodology and Results
A statistical-based deformable model is being developed that improves upon existing point distribution models (PDMs). As mentioned above, existing PDM boundary finding techniques often suffer from the following shortcomings: (1) a priori local shape characteristics are not utilized, (2) global shape and gray-level information are treated independently during boundary optimization, and (3) there is no existing metric that provides a confidence measure of segmentation performance. A new deformable model algorithm is under development in which the objective function used during optimization of the boundary encompasses several important characteristics. First the objective function includes both global shape and local gray-level characteristics, so optimization occurs with respect to both pieces of information simultaneously. In addition, local shape characteristics, as derived from the training set, are also incorporated into the boundary finding process. Finally, the objective function is formulated in a way that leads to a confidence metric that indicates how well the final boundary fits the underlying object as defined in the target image. This new algorithm is being applied to geometric test images as well as high-resolution x-ray computed tomography (CT) images of laboratory mice for the purpose of organ identification.

This work is being conducted at IRIS lab by Shaun S. Gleason under the supervision of Dr. M. A. Abidi. The research is sponsored by the Laboratory-Directed Research and Development Program of ORNL under contract DE-AC05-96OR22464.