Nonlinear Dimensionality Reduction of Brain MRI Data for Classification Applications

点击翻译按钮获取中文摘要

Classification, or the categorization of images into meaningful groups, is an important application in brain image analysis. Dimensionality reduction is a key component of many classification algorithms and is essential for reducing the complexity of mathematical methods for working with brain images. A simple example of dimensionality reduction in everyday use is the projection of 3D geography onto 2D maps. Different projection methods preserve different properties of the geography, just like different dimensionality reduction methods preserve different properties of brain images. Recently, groundbreaking discoveries in the field of machine learning have rekindled a strong interest in a type of neural network called the deep belief net (DBN), which has been shown to be very effective at automatically learning patterns of similarity in large groups of small images by reducing the dimensionality of the images to form a simplified virtual landscape in which similar images occupy the same valleys and dissimilar images are separated by hills. This approach has been shown to be more accurate for identifying faces and handwritten digits than practically all other methods. In comparison to the previous data used for DBN experiments, brain images are of much higher dimension and have much more complex shape and intensity variations. Our long-term research goal is to develop new DBN methods that can be used to form simplified landscapes of groups of brain images such that the valleys represent meaningful similarities that can be used for classification. Using a very large database of magnetic resonance images (MRIs) of patients with multiple sclerosis, we will develop new methods for making DBNs work with large, complex images. We will investigate what types of similarities DBNs are capable of extracting from groups of brain images, and determine whether those similarities are potentially useful for classifying brain images into various types of categories that will range from the visually obvious to ones that are not image-derived, such as clinical outcomes. If successful, this work will greatly benefit researchers in neuroimaging by allowing them to make much more effective use of their data.