Alzheimer’s disease diagnosis on structural MR Images using Circular Harmonic Functions descriptors on Hippocampus and Posterior Cingulate Cortex
Résumé
Recently, several pattern recognition methods have been proposed to automatically
discriminate between patients with and without Alzheimer’s disease using different imaging
techniques: sMRI, fMRI, PET and SPECT. Classical approaches in visual information
retrieval have been successfully used for analysis of structural MRI brain images. In this
paper, we use the visual indexing framework and pattern recognition analysis based on
structural MRI data to discriminate three classes of subjects: Normal Controls (NC),
Mild Cognitive Impairment (MCI) and Alzheimer’s Disease (AD). The approach uses the
Circular Harmonic Functions (CHFs) to extract local features from the most involved
areas in the disease: Hippocampus and Posterior Cingulate Cortex (PCC) in each slice
in all three brain projections. The features are quantized using the Bag-of-Visual-Words
approach to build one signature by brain (subject). This yields a transformation of a full
3D image of brain ROIs into a 1D signature, a histogram of quantized features. To reduce
the dimensionality of the signature, we use the PCA technique. Support vector machines
classifiers are then applied to classify groups. The experiments were conducted on a subset
of ADNI dataset and applied to the ”Bordeaux-3City” dataset. The results showed that
our approach achieves respectively for ADNI dataset and ”Bordeaux-3City” dataset; for AD vs NC classification, an accuracy of 83.77% and 78%, a specificity of 88.2% and 80.4% and a sensitivity of 78.54% and 74.7%. For NC versus MCI classification we achieved
for the ADNI datasets an accuracy of 69.45%, a specificity of 74.8% and a sensitivity of
62.52%. For the most challenging classification task (AD versus MCI), we reached an
accuracy of 62.07%, a specificity of 75.15% and a sensitivity of 49.02%. The use of PCC visual
features description improves classification results by more than 5% compared to the use
of Hippocampus features only. Our approach is automatic, less time-consuming and does
not require the intervention of the clinician during the disease diagnosis.