A Novel Scheme for the Classification Analysis of Big Image Data Based on Functional Principle Component Analysis, Matrix Completion, and Sufficient Dimension Reduction

• Activity Number: 655
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Nan Lin*+ and Junhai Jiang and Shicheng Guo and Xiao Yu and Long Ma and Momiao Xiong
• Companies: UTSPH and UTSPH and University of Texas and University of Texas and University of Texas and University of Texas Health Science Center at Houston
• Keywords: image analysis ; functional data analysis ; matrix completion ; sufficient dimension reduction

Abstract

In this talk, we address three key issues for image analysis: data representation, data denoising and feature selection. (1) We extend one dimensional principle component analysis to two dimensional principle component analysis (2DFPCA) and use 2DFPCA scores to reduce the dimensions of the image data in the frequency domain. (2) Instead of applying widely used image denoising methods that can only effectively deal with a single type of noise, we use recently developed matrix completion methods to remove mixture noise from the image data. (3) The most traditional feature selection methods without taking information on class labels or disease status into account seriously compromise the image diagnostic accuracy. To overcome this limitation, we develop a novel sparse sufficient dimension method that can substantially reduce the number of features while keeping class or disease status information in the reduced features. The proposed methods are applied to 365 liver cancer ultrasound images. By 5-fold cross validation, the average classification accuracy, sensitivity and specificity in the test set are 72%, 68% and 76%, and in the training set are 81%, 77% and 85%, respectively.