Multimodality Imaging Population Analysis using Manifold Learning

International audience ; Characterizing the variations in anatomy and tissue properties in large populations is a challenging problem in medical imaging. Various statistical analysis, dimension reduction and clustering techniques have been developed to reach this goal. These techniques can provide insight into the effects of demographic and genetic factors on disease progression. They can also be used to improve the accuracy and remove biases in various image segmentation and registration algorithms. In this paper we explore the potential of some non linear dimensionality reduction (NLDR) techniques to establish simple imaging indicators of ageing and Alzheimers Disease (AD) on a large population of multimodality brain images (Magnetic Resonance Imaging (MRI) and PiB Positron Emission Tomography (PET)) composed of 218 patients including healthy control, mild cognitive impairment and AD. Using T1-weighted MR images, we found using laplacian eigenmaps that the main variation across this population was the size of the ventricles. For the grey matter signal in PiB PET images, we built manifolds that showed transition from low to high PiB retention. The combination of the two modalities generated a manifold with different areas that corresponded to different ventricle sizes and beta-amyloid loads.