Content-based indexing for medical image databases

In a large medical image databases system, a content-based indexing structure is often established from the image feature vectors so as to allow fast retrievals of medical images. However, these vectors will generally be having a high number of dimensions which will then result in poor indexing performance. In this paper, we investigate how to improve the search performance of the packed R-tree when its indices are of high dimensions. Two new algorithms are designed according to their different approaches in applying the idea of principal component analysis (PCA) technique. The first algorithm performs a dominant dimension analysis globally, and selects the first few dominant dimensions for the packing steps. Further, the same set of dominant dimensions are used in calculating image similarities afterwards. The second algorithm is differed from the first one by re-applying the analysis at each tree node, and hence obtaining a better set of dominant dimensions of the image data under the sub-tree headed by the node. In developing the second algorithm, we have also considered how to reduce the calculations by utilizing the results of the tree nodes at lower levels. This paper reports the performance of the two algorithms with different data sets. The algorithms are tested with a set of random generated images, and a real medical image database of about 2,000 MRI. In the experiments, we observe a better retrieval performance in the second algorithm. Similar results are reported even when the data are highly randomized.