Volume visualization and exploration through flexible transfer function design

Abstract

Direct volume rendering (DVR) is a well-known method for exploring volumetric data sets. Optical properties are assigned to the volume data and then a DVR algorithm produces visualizations by sampling volume elements and projecting them into the image plane. The mapping from voxel values to optical attribute values is known as transfer function (TF). Therefore, the quality of a visualization is highly dependent on the TF employed, but its specification is a non-trivial and unintuitive task. Without any help during the TF design process, the user goes through a frustrating and time-consuming trial-and-error cycle. This paper presents a useful combination of TF design techniques in an interactive workspace for volume visualization. Our strategy relies on semi-automatic TFs generation methods: boundary emphasis, stochastic evolutive search in TF space, and manual TF specification aided by dual domain interaction. A two-level user interface was also developed. In the first level, it provides multiple simultaneous interactive visualizations of the volume data using different TFs, while in the second one, a detailed visualization of a single TF and the respective rendered volume are displayed. Moreover, in the second level, the TF can be manually refined and the volume can be further inspected through geometric tools. The techniques combined in this work are complementary, allowing easy and fast TF design and data exploration.

Introduction

Volume rendering is widely known as a set of methods for visualization of large three-dimensional (3D) scalar or vector fields, mainly in medical and scientific data exploration. In these areas, one often deals with 3D images, such as those obtained from CT and MRI devices or numerical simulation data. Volume rendering techniques and algorithms are well described in the literature [1], and can be classified as isosurface extraction methods and direct volume rendering (DVR) methods. The former methods extract polygonal meshes representing isosurfaces in the volume and then use the traditional rendering pipeline to display the meshes (see [2], for a well-known example). On the other hand, DVR methods display volume data without extracting an intermediate geometry. DVR and its advantages were first described by Levoy [3]. Modern graphics hardware supports volume rendering at interactive rates using either of these approaches.

To obtain useful images through DVR, voxels have to be classified in order to determine which ones must be displayed. This classification is typically performed by transfer functions (TFs), which associate values of optical attributes to voxels based on their values. Opacity and color are the most common optical properties used in TFs. The degree of opacity can make a voxel more or less visible and is normally used to emphasize voxels in boundaries between different homogeneous regions of the volume [4]. Other optical properties may also be used, such as specular reflection coefficients [5], spectral reflectance [6] and light scattering coefficients [7]. More recently, the concept of style TFs was introduced by Bruckner and Gröller [8], who used TFs to define the rendering style of volume regions based on data values and eye-space normals. The information conveyed by an image built from volume data is, therefore, highly dependent on the quality of the TF. However, TF design is a non-trivial and unintuitive task, and has been referred as one of the top 10 problems in volume visualization [9].

One-dimensional (1D) TFs take into account only scalar voxel values, and are the most common TFs, although having a limited classification power. On the other hand, multi-dimensional TFs allow more freedom in voxel classification by taking as arguments vectorial values or combinations of local measures of scalar fields, such as derivative values [10], [11], neighborhood, position [12], curvature [13], [14] and statistical signatures [15]. Notwithstanding, the design complexity grows with the size of the TF domain [10], and the memory required to implement truly multi-dimensional TFs restricts their application [16]. In this work, we adopted 1D TFs due to their simplicity and low-memory requirements, since they can be implemented as small lookup tables. Furthermore, the pre-integrated volume rendering technique, proposed by Engel et al. [17], allows high-quality DVR at interactive rates using 1D TFs.

Designing TFs with no assistance leads to trial-and-error efforts. Therefore, several automatic and semi-automatic techniques for specifying TFs have been proposed [4], [8], [15], [18], [19], [20], [21], [22], [23]. They can be guided by analysis of the volumetric data (data-driven) or analysis of the generated images (image-driven) [9]. In any case, to make the process less frustrating and less time-consuming, the user must be given a rapid feedback with real-time rendering frame rates.

The main contribution of our work is a two-level interface method that combines a set of useful tools for semi-automatic 1D TF design and fast data exploration in an interactive workspace. The first level of our interface presents several thumbnails of the volume data rendered with different TFs, allowing immediate insight of the main structures in the data set. The second level shows a detailed visualization of a single TF as well as the resulting rendering. TFs can be easily generated and refined using semi-automatic boundary emphasis, stochastic evolutive search and manual design aided by dual domain interaction [11]. These three approaches are complementary and were successfully combined in this work, improving the idea of two-level interaction proposed by Prauchner et al. [24].

This paper extends our previous work [25] with an improved interface and an experimental evaluation of our approach. We implemented a history tree that keeps track of the TF evolution and allows the user to go back to a previously specified TF. The evaluation was performed as an experiment with 15 subjects who accomplished two visualization tasks with different data sets. This way we tested the usability of our methods with potential users. We also implemented a set of geometric tools for volume inspection inspired by the work of Dietrich et al. [26], but their description is beyond the scope of this article.

The paper is organized as follows. Section 2 discusses the closest related works. Section 3 describes the proposed interface and the TF design techniques provided within it. Implementation details are addressed in Section 4, while in Section 5 we present the evaluation of our tools. At last, in Section 6, we draw some conclusions and point directions for future work.