A COMPUTATIONAL FRAMEWORK FOR DISCRIMINATIVE ANALYSIS OF HIGH DIMENSIONAL BIOMEDICAL IMAGE DATA

In this work we propose, implement and evaluate a discriminative computational pipeline to address the problem of feature generation, feature screening and feature subset selection from biomedical datasets when the ratio of feature dimensions (possibly in millions) to number of samples is very high . The proposed pipeline is modular and can be highly parallel. The framework is applied to a variety of real world discrimination problems including plant specie classification from 2D leaf images, Gender/age/expression/human identification from 3D facial surface mesh and Gender/Age/Disease classification from 3D neural Magnetic Resonance Imaging (MRI) images. By using a unique set of novel features for each application we either achieve better or competitive classification accuracy than existing work or set new benchmarks otherwise. We can also locate discriminative regions for faces and brains which facilitate further scientific discoveries. This work illustrates quantitatively, the effectiveness and the diversity of the proposed feature extraction and machine learning pipelines.