Computationally intelligent methods for mining 3D medical images

Abstract

We present novel intelligent tools for mining 3D medical images. We focus on detecting discriminative Regions of Interest (ROIs) and mining associations between their spatial distribution and other clinical assessment. To identify these highly informative regions, we propose utilizing statistical tests to selectively partition the 3D space into a number of hyper-rectangles. We apply quantitative characterization techniques to extract k-dimensional signatures from the highly discriminative ROIs. Finally, we use neural networks for classification. As a case study, we analyze an fMRI dataset obtained from a study on Alzheimer's disease. We seek to discover brain activation regions that discriminate controls from patients. The overall classification based on activation patterns in these areas exceeded 90% with nearly 100% accuracy on patients, outperforming the naïve static partitioning approach. The proposed intelligent tools have great potential for revealing relationships between ROIs in medical images and other clinical variables assisting systems that support medical diagnosis.