Robust Multi-Landmark Detection Based on Information Theoretic Scheduling

Abstract An anatomical landmark is a biologically meaningful point of an organism. In the medical imaging domain, anatomical landmarks provide guidance of image navigation and evidence for anomaly diagnosis. Therefore, anatomical landmark detection becomes one of the fundamental tasks in various medical image analysis systems. In most of these systems, multiple landmarks are required to be detected fast and robustly. In this chapter, we present a method for robust and efficient multi-landmark detection based on information theory. While the detection of each landmark is considered as a “vision task” of a medical image analysis system, our method aims to schedule these vision tasks optimally in an information-theoretic sense. The key idea is to schedule tasks in such an order that each task/operation achieves maximum expected information gain over all the tasks. The scheduling rule is formulated to embed two intuitive principles: (1) a task with higher confidence tends to be scheduled earlier and (2) a task with higher predictive power for other tasks tends to be scheduled earlier. More specifically, task dependency is modeled by conditional probability; the outcome of each task is assumed to be probabilistic as well; and the scheduling criterion is based on the reduction of the summed conditional entropy over all tasks. We demonstrate our algorithm on six diverse applications relevant to different imaging modalities (computed tomography, magnetic resonance, positron emission tomography, and radiograph) and clinical fields (neurology, oncology, and orthopedics).