Landmark detection in cardiac MRI using learned local image statistics

Abstract

We propose a supervised learning approach for detecting landmarks in cardiac images from different views. A set of candidate landmark points are obtained using morphological operations and graph cut segmentation. The final landmarks are determined using random forests (RF) classifiers which were trained on low level features derived from the neighborhood of annotated landmarks on training images. We use features like intensity, texture, shape asymmetry and context information for landmark detection. Experimental results on the STACOM LV landmark detection challenge dataset show that our approaching is promising with room for further improvement. © 2013 Springer-Verlag.