Discriminative Joint Context for Automatic Landmark Set Detection from a Single Cardiac MR Long Axis Slice

Cardiac magnetic resonance (MR) imaging has advanced to become a powerful diagnostic tool in clinical practice. Automatic detection of anatomic landmarks from MR images is important for structural and functional analysis of the heart. Learning-based object detection methods have demonstrated their capabilities to handle large variations of the object by exploring a local region, context, around the target. Conventional context is associated with each individual landmark to encode local shape and appearance evidence. We extend this concept to a landmark set, where multiple landmarks have connections at the semantic level, e.g., landmarks belonging to the same anatomy. We propose a joint context approach to construct contextual regions between landmarks. A discriminative model is learned to utilize inter-landmark features for landmark set detection as an entirety. This helps resolve ambiguities of individual landmark detection results. A probabilistic boosting tree is used to learn a discriminative model based on contextual features. We adopt a marginal space learning strategy to efficiently learn and search a high dimensional parameter space. A fully automatic system is developed to detect the set of three landmarks of the left ventricle, the apex and the two basal annulus points, from a single cardiac MR long axis image. We test the proposed approach on a database of 795 long axis images from 116 patients. A 4-fold cross validation results show that about 15% reduction of the errors is obtained by integrating joint context into a conventional landmark detection system.