Integrating discretization and association rule-based classification for Alzheimer’s disease diagnosis

This paper shows a computer aided diagnosis (CAD) combining continuous attribute discretization and association rule mining for the early diagnosis of Alzheimer’s disease (AD) based on emission computed tomography images. A mask is obtained from the mean control images by an image histogram segmentation. 3D voxels centered in mask coordinates are selected by equal-width binning-based discretization of the mean intensity. These Regions of Interest (ROIs) are then used as input for the Association Rule (AR)-mining using control subject images to fully characterize the normal pattern of the image. Minimum support and confidence are fixed to the maximum values in order to obtain the highest predictive power rules for each discretization level (or combination of levels). Finally, classification is carried out by comparing the number of ARs verified by each subject under test. The proposed system is evaluated using two different databases of single photon emission computed tomography (SPECT) and positron emission tomography (PET) images from the Alzheimer Disease Neuroimaging Initiative (ADNI) yielding an accuracy up to 96.91% (for SPECT) and 92% (for PET), thus outperforming the baseline (the so called continuous AR-based method) and other recently reported CAD methods.

► Computer aided diagnosis (CAD) system design for the early diagnosis of Alzheimer’s Disease (AD).
► Image histogram segmentation for masking and equal-width binning discretization of continuous attributes for ROI selection.
► AR discovery among discrete ROIs from controls.
► Tested on a 97-SPECT image database and 150-PET image ADNI database.
► Accuracy rates obtained by an AR-based classification method are over 90% for SPECT and PET images.