Brain MRI landmark identification and detection

Abstract

Knowledge of the location of anatomical landmarks on the brain is important in neuroimaging. Applications include landmark-based image registration, segmentation of brain structures, electrode placement in deep brain stimulation, and prospective subject positioning in longitudinal imaging. Landmarks are specific structures with distinguishable morphological characteristics. In this study, we only consider point landmarks on magnetic resonance imaging (MRI) brain scans. The most basic method for locating anatomical landmarks on MRI is manual placement by a trained operator. However, manual landmark detection is a strenuous and tedious task, especially if large databases are involved and/or multiple landmarks need to be located. Therefore, automatic landmark detection on MRI has become an active area of research. Model-based methods are popular for detecting brain landmarks. Generally, model-based landmark detection includes a training set of MRI scans on which the location of certain landmarks are known, usually by manual placement. The location of landmarks on the training set is then used to derive and store models for individual landmarks. Then, when the same landmarks are to be located on a test MRI volume, the models are recalled and their information is used to automatically detect the landmarks. In this thesis, we propose a new unsupervised landmark identification method for the training phase of this process to replace manual landmark identification on the training set of MRI volumes. This method employs an iterative algorithm for detecting a set of landmarks on the training set that are leave-one-out consistent. In addition, we suggest a detection method to locate the corresponding points on a given test volume. In this study, the method was implemented and applied to a dataset of sixty 3D MRI volumes. The training was performed on 30 volumes. The remaining 30 volumes were used as a test set on which the detection algorithm located the corresponding landmarks. In the landmark identification approach, a set of candidate seeds are necessary as the initial guesses of landmark positions. The position and number of the seeds are optional. In this study, we used 154 candidate seeds spread uniformly across the entire brain volume. All the identified and detected landmarks were inspected manually us- ing a graphical user interface. To further evaluate the performance of the introduced method, we registered a set of 152 brain images to a reference space employing this method. Brain overlap of the registered volumes improved as a result of landmark based registration. As a further application, we used landmark detection for rigid-body registration of longitudinal MRI volumes. These are MRI volumes scanned from the same individual over time. We show that landmark detection is a fast method that can be used to obtain a good initial rigid-body registration which can then be followed by fine-tuning of the registration parameters.