Landmark localization on color coded diffusion anisotropy images using convolutional neural networks

Landmark localization, finding exact location of structures in an image is a first stage of many complex computer vision problems. Locating specific landmarks on brain images is one of the stages in defining the target in functional surgery and in estimating point wise correspondence in image registration. Nowadays, various types of convolutional neural networks (CNN) have been proposed that are able to interpret complex computer vision problems. In this study, a CNN based landmark detector is employed to locate specific landmarks at given MNI coordinates, on an individual's diffusion MR brain images. MR diffusion images, with their high degree of heterogeneity, especially in white matter, provide a rich set of features compared to other basic structural images such as T1 or T2 weighted images. Results show that finding a specific point on brain using diffusion characteristics by CNN based model is sustainable and has a potential to be a base for image registration techniques.