Anatomical landmark detection leveraging implicit and explicit information sharing techniques

Anatomical landmarks are used to position human brain images for providing a reference coordinate system. Having a consistent brain alignment between scanning sessions and subjects enables researchers to track structural changes in the brain over time and make differential comparisons between subjects. Anatomical landmark detection has traditionally been solved by leveraging known characteristics about the landmarks and known relationships between landmarks. These systems are sensitive to missing data and are specific to the anatomy that is scanned.

Recent work has shown that landmark detection can be generalized and performed with reasonable accuracy by treating the detection process as an intelligent search through the image performed by an agent. In this work, we build upon this intelligent search method in two ways. First, we search for a few high confidence landmarks in the original unaligned scan acquisition space. Second, the majority of landmarks to be detected with the image in a consistent orientation proscribed by the first stage high confidence landmarks. This consistent orientation enables more effective agent initial placement in the image to start their search and creates a more consistent relationship between landmarks.