Segmentation by multispectral analysis of mr image data for 3d modelling in surgical planning

Magnetic resonance imaging is the method of choice for the examination of extremities and its joints. Usually a study comprises several image series acquired with different pulse sequences, providing a broad variation of tissue contrast. Acceptable signal-to-noise ratio is the reason for higher slice thickness compared to CT, yielding rougher 3D models for further surgical planning. In this work a newly developed image processing chain is presented exploiting the multispectral nature of standard MRI studies for accurate segmentation of the shoulder joint. After careful image registration, employing mutual information, Gaussian clustering is applied for the identification of relevant tissue types. The algorithm is potentially self-learning, but needs some post processing to yield lifelike structures. Segmented objects are smoothed by binary morphological operations and transformed to isotropic resolution by shape based interpolation. Results of each step of the processing pipeline are presented. The usefulness of standard clinical MRI studies for the generation of accurate 3D models is demonstrated with a showcase for surgical planning.