A Comparative Study of Automatic Localization Algorithms for Spherical Markers within 3D MRI Data.

Localization of features and structures in images is an important task in medical image-processing. Characteristic structures and features are used in diagnostics and surgery planning for spatial adjustments of the volumetric data, including image registration or localization of bone-anchors and fiducials. Since this task is highly recurrent, a fast, reliable and automated approach without human interaction and parameter adjustment is of high interest. In this paper we propose and compare four image processing pipelines, including algorithms for automatic detection and localization of spherical features within 3D MRI data. We developed a convolution based method as well as algorithms based on connected-components labeling and analysis and the circular Hough-transform. A blob detection related approach, analyzing the Hessian determinant, was examined. Furthermore, we introduce a novel spherical MRI-marker design. In combination with the proposed algorithms and pipelines, this allows the detection and spatial localization, including the direction, of fiducials and bone-anchors.

Evaluation of electroencephalography analysis methods.

The extraction of expressive features from an electroencephalography (EEG) signal is necessary for classification of movement and movement imagination of the limbs. We introduce different preprocessing and feature extraction algorithms for this purpose and develop an algorithm that selects features by their feature importance. This selection is used as an evaluation measure for features, their preprocessing algorithms and the EEG electrodes. Our results show that most influential features for signal interpretation are: common spatial patterns, fractal dimensions, as well as, variance and standard deviation of the preprocessed data. We show that preprocessing with continuous wavelet transforms outperforms the other tested preprocessing algorithms. Furthermore, we show that high gamma frequencies (70-90 Hz) contain more information than the lower µ-rhythms (8-12 Hz) where event-related-desynchronization (ERD) is known to occur. The important EEG electrodes for this classification task are located in the left and right back of the motor-cortex. The proposed algorithm can be further used to create subject-specific and performance models for real-time classification.