<title>Generalized transfer functions and their use for three-dimensional neuroimaging</title>

Abstract

JHU-ISS Center for Information-enhanced MedicineInstitute of Systems Science, National University of SingaporeHeng Mui Keng Terrace, Kent Ridge, Singapore 0511ABSTRACTVolume rendering uses classification based on fuzzy segmentation, and the key to making meaningful images lies in aproper choice of the opacity and color transfer functions. The smooth and continuous transfer functions introduce less artifactsthan any binary operations such as thresholding which disrupts the continuity of the data. Despite this advantage, densitybased fuzzy segmentation has still several limitations. To find suitable transfer functions for real clinical data may be a labori-ous task, and methods facilitating the automated generation of transfer functions are very useful. Furthermore, the standardtransfer functions are based on the density of a resampled point. This results in several shortcomings, e.g., the inability to dis-play clearly a non-homogeneous object such as a tumor which density range overlaps with those of its surrounding structures.Moreover for MRI head data, the same density range may correspond to different structures which are then mapped into thesame color/grayscale. To overcome those limitations, we propose a suitable extension of the standard transfer functions calledgeneralized transferfunctions. These functions use both density based as well as non-density based information about classi-fled structures, such as a contour information about the tumor, space-dependent information, multi-modal information, higherorder information, anatomical knowledge. We show the usefulness of the generalized transfer functions in 3D neuroimaging(neuropathology) from MRI data. Three approaches are discussed: contour-enhanced volume rendering, ROIs-enhanced vol-ume rendering, and slice density corrected transfer functions.1. INTRODUCTIONThree-dimensional visualization techniques are considered very useful for computer assisted surgery, radiation therapy,complex fractures or abnormalities, and many others. These techniques are based on a large variety of principles, such asvoxel or polygon data structures, binary or fuzzy surface segmentation, opaque or transparent rendering, surface normal cal-culation from depth buffer, gray level gradients or polygon inclinations, and the use of various illumination models. The mostpopular techniques are surface rendering1'2 and volume rendering3'4'5'6'7. Surface rendering uses a surface detector to thebinary scene, fit geometric primitives (polygons) to the detected surfaces, and render the resulting geometrical representation.To transform the 3D array of voxels into the binary scene, thresholding is used. Volume rendering does not explicitly extractsurfaces, but instead it uses all the voxels of the volumetric dataset. Rendering the dataset involves resampling the dataset andmapping each resampled value into a color and opacity. The opacities and colors are composited with each other on depth-sorted order yielding a color of the pixel. The main advantages of volume rendering over other visualization techniques are itssuperior image quality and ability to generate images without explicitly defining surfaces representing boundaries between