Organ Labeling Using Anatomical Model-Driven Global Optimization

Organ labeling involves estimating approximate organ locations from medical images and has several traditional and emerging applications. The organ labels can serve as initialization for automatic and fast image processing, provide additional information for image data mining and retrieval, and assist fast image navigation and workflow on radiology workstations. The problem is made difficult by variability in anatomy, imaging modalities, scanning protocols, and pathology. This paper presents a new model-based framework to localizing multiple organs in the body from images produced by different scanning modalities. The model has two components: a modality-independent geometric model and a modality-specific (CT, MR) appearance model. The geometric representation encodes anatomical knowledge as the distribution of individual and relative locations of organs. The appearance models for each organ are created from their intensity profiles in the respective modality. The multi-organ labeling problem is then formulated as a global optimization problem solved using dynamic programming without requiring initialization. Initial results of applying the framework to CT and MR volumes were encouraging and demonstrated the strength of using organ-level anatomical information with simple geometric models and image-level feature sets for the task of organ labeling.