Analysis of Echocardiography Images using Manifold Learning

Dimensionality reduction is an important task in machine learning, to simplify data mining, image processing, classification and visualization of high-dimensional data by mitigating undesired properties of high-dimensional spaces. Manifold learning is a relatively new approach to nonlinear dimensionality reduction. Algorithms for manifold learning are based on the intuition that the dimensionality of many data sets may be artificially high and each data point can be described as a function of only a few underlying parameters. Using this tool, intrinsic parameters of the system database, which are main distinction factors of data sets, are recognized and all of them lie on a manifold that shows the real relationship of parameters. One of the successful applications of these methods is in image analysis field. By this approach, each image is a data in high dimensional space that the pixels are its dimensions. Because echocardiography images obtained from a patient are different in quantitative parameters such as heartbeat periodic motion and noise, image sets are reduced to two-dimensional space by a proper manifold learning. In this article, after mapping echocardiography images in two-dimensional space, by using LLE and Isomap algorithms, similar images placed side by side and the relationships between the images according to the cyclic property of heartbeat became evident. The Results showed the weakness of Isomap algorithm and power of LLE algorithm in preserving the relation between consecutive frames. De-noising is an important application which extracted from this research.