Chapter 4 - Landmark Detection Using Submodular Functions

Abstract

We present a method for detecting a large number of anatomical structures, or landmarks, in a provably efficient way. A trivial solution is to search for each landmark independently. This is computationally expensive and cannot scale to a large number of landmarks. We devise a greedy algorithm that detects landmarks one by one, with the goal of minimizing the total computational cost. While this objective function is provably hard to solve, we prove that the greedy algorithm achieves near optimal results. The system is validated on a database of 2500 computed tomography volumes and shows significant speedup over traditional methods.