Content-based Retrieval of 3D Medical Images

Abstract -- While content-based image retrieval (CBIR) has been researched for more than two decades, retrieving 3D datasets has been progressing considerably more slowly, especially in respect to its application to the medical domain. This is in part due to the limitation of processing speed when trying to retrieve high-resolution datasets in real-time. Another barrier is that most existing methods have been developed based on 2D images instead of 3D, leaving a gap to be filled. At present, a significant number of exploitations are focusing on the extraction of 3D shapes. However, it appears other information tends to be equally important in clinical decision making. In this paper, Local Binary Pattern (LBP), a texture based approach stemming from 2D forms, has been studied extensively through the application to 3D images from a collection of MR brain images in a content-based image retrieval system (CBIR). The initial results show LBP not only can achieve a precision rate of up to 78% but also can perform retrieval in real time with sub-second processing speeds. Comparison with the other three popular texture-based methods, namely 3D Grey Level Co-occurrence Matrices, 3D Wavelet Transforms and 3D Gabor Transforms, is also carried out. The results demonstrate that LBP outperforms them all in terms of retrieval precision and processing speed. Keywords – CBIR, 3D image retrieval, 3D texture extraction.