Section: New Results

Evaluation of morphometric descriptors of deep brain structures for the automatic classification of patients with Alzheimer's disease, mild cognitive impairment and elderly controls

Participants : Alexandre Routier [correspondant] , Pietro Gori, Ana Fouquier, Sophie Lecomte, Olivier Colliot, Stanley Durrleman.

We participated in the Computer-Aided Diagnosis of Dementia based on structural MRI data (http://caddementia.grand-challenge.org/ ). Our approach was to select shapes of 12 brain structures: the caudate nucleus, putamen, pallidum, thalamus, hippocampus and amygdala of each hemisphere. The structure segmentation was based on a FreeSurfer segmentation and the marching-cubes algorithm was used to get 3D triangular meshes. Using our software Deformetrica, anatomical models (mean shape and typical variations) of these brain structures were built for patients with Alzheimer's disease (AD), Mild Cognitive Impairments (MCI) and cognitively normal controls (CN) based on the data of 509 ADNI subjects. The models for AD, MCI and CN were registered to the test subjects by maximizing the likelihood of the test image to be derived from each model. The final classification was made by thresholding this criterion taking into account the covariance of the deformation parameters. The thresholds were either optimized on the ADNI data or on the provided training data. The method was fully automatic and the computation time was 4 days for training the anatomical models plus 11 hours per subject for registration and classification. For the 30 training subjects, the algorithm had accuracies of 73% (if optimized on training data) and 50% (if optimized on ADNI data). On the test set of 354 images, our method yields an accuracy of 49.2% (43.5 - 54.2), true positive fraction of 94.6% (89.8 - 97.7) for CN, 11.5% (6.2 - 17.7) for MCI and 36.9% (27.4 - 46.5) for AD.

Our participation to this challenge was the opportunity to test our software Deformetrica for classification tasks. It ran on more than 800 images, thus showing its ability to deal with large data sets.

More details in [27] .