Purchase this article with an account.
P. Sethi, S. Dua, R.W. Beuerman, M. Hartnett; Real-Time Indexing of Retinal Images for Data Mining and Content-Based Image Retrieval Applications . Invest. Ophthalmol. Vis. Sci. 2003;44(13):3652.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: To propose and demonstrate a unique algorithm for real-time, fast indexing of retinal images for data mining applications. In clinical imaging, the speed of data collection has surpassed our ability for rapid computational analysis. A successful data mining technique to discover embedded patterns or anomalies in retinal image databases depends on the degree of organization (indexing) of this high-dimensional data to enable similitude search. It has been shown that any fast computational technique behaves exponentially in performance and space with the increasing dimensionality. Hence, development of unique indexing schemas for retinal image data enabling similarity based search for data mining applications and content based image retrieval for computer-aided analysis and diagnosis, is important. Methods: Our experimental and analytical analysis has shown that most of the energy in retinal images is concentrated in the first few Fourier (orthonormal) coefficients. We exploit this property to uniquely represent an image as a 1-D signal in time domain, dividing it in overlapping windows of fewer Fourier features. These windows are then identified as points in a unique feature trail in Fourier space. These feature points are clustered having relatively stationary harmonic behavior, identified by their convex hulls and approximated by their minimum enclosing rectangles (MER). These MERs are then arranged in a hierarchical, multi-dimensional index for similitude search. Results: Proposed similarity-metric is shown to be preserved in orthonormal transformations and the index is demonstrated to be rotational, translational and scale-invariant. The proposed index is demonstrated to be independent of the size of images and is 70% faster in search-speed than commonly used sequential search. The proposed index is also shown to have extremely low (< 1%) false dismissals and few (< 5%) false alarms. This new technique also exhibits 27% reduction in query processing time over previous approach which was only reported for a database of equal-sized (image) sequences. In contrast, this proposed index can also report similar unequal sized images. Conclusions: The proposed index is translational, rotational and scale in-variant and far exceeds sequential search in search performance. We suggest this approach is useful for synthetic and real data, propose applications and define avenues for further refinements.
This PDF is available to Subscribers Only