Visualization Of Photographic Volumes Using Machine Learning

In recent years, the visualization technology plays an increasingly important role in the medical field. The continuous scanned medical images can represent more anatomical and pathological information by 3d reconstruction. Photographic volumes, which are derived from physiological slice of human beings, provide more data information than traditional scalar volumes, as the realistic color of each sampled position is recorded in the three-dimensional vector of each voxel. However, due to the lack of physical information, such as density value and absorption coefficient, photographic volumes produce a different challenge for the classification, especially the specification of opacity. Effective and reasonable classification of volume data and appropriately setting of opacity value to the corresponding voxel can help to extract features of interest. Many techniques have been developed for the effective classification of scale volumes, while the research on classification of photographic volumes is limited.Aiming at the above-mentioned issues, this paper analyzes the properties of photographic volumes, such as color information and derived texture value, and presents a novel clustering based classification for photographic volumes, which is inspired by machine learning.Gaussian mixture model (GMM) is first proposed for the initiative classification of photographic volume in the color space. The user can further adjust the parameters of each Guassian model in order to achieve a better classification result.Texture analysis, which is a fundamental technique in image processing, is employed to analyze the characteristic patterns and textures of photographic volumes, which are difficult to identity in the color space. We can use a combination of first-, second-and high-order statistics to capture textural properties and then apply the Local Linear Embedding (LLE) technique to map the high-dimensional textural properties into a global low-dimensional feature space. In this way, features of interest can be identified intuitively through the interactive window selection on the reduced feature space.The visualization results of various photographic volumes demonstrate that our proposed classification methods, which are based on GMM and LLE, are very effective in the classification of photographic volumes.