I joined the Aramis Lab (Inria, CNRS UMR 7225, Inserm U1127, ICM, Sorbonne Université) in January 2017 after having been awarded a PRESTIGE postdoctoral research fellowship (a Marie Skłodowska-Curie fellowship programme) to develop computational imaging tools to improve the understanding and diagnosis of dementia.
Dementia affects over 7 million people in Western Europe only. There is currently no cure that can slow down or stop the course of these diseases, and the development of new treatments is hampered by the difficulty in identifying the correct type of dementia at an early stage. This difficulty is partly due to the fact that a large amount of information, such as multiple images of the brain, needs to be analysed before making a diagnosis. The goal of this project is to automatically locate and characterise the typical areas of different dementia stages and syndromes by jointly exploiting the structural and functional information provided during MRI and PET examinations. This project will lead to the definition of abnormality maps summarising the pathology’s topographical distribution in the brain. The estimated abnormality information will help distinguish between dementia subtypes by providing both qualitative and quantitative information. It will i) help clinicians in their diagnosis by highlighting, in a data-driven fashion, the pathological areas obtained from the individual PET and MRI data, and ii) provide quantitative, voxel-based, abnormality scores that can be used as input for computer-assisted diagnosis (CAD) tools for the automatic discrimination between dementia stages and subtypes.
My first objective has been to develop a strategy to reduce the confounding impact of anatomical variability when using machine learning to distinguish disease versus normal ageing. In machine learning classification methods developed for dementia studies, neuroimaging features, e.g. glucose consumption extracted from PET images, are often used to draw the border that differentiates normality from abnormality. However, these features are affected by the anatomical and metabolic variabilities present in the population, which act as a confounding factor making the task of finding the frontier (i.e. the decision function) between normality and abnormality very challenging. Instead of trying to find this frontier at the population level, I proposed to transport the problem to the individual level to reduce its complexity. During my PhD I developed a method able to extract for each individual the signal characteristic of abnormality from PET data. This framework consists of creating a patient-specific model of healthy PET appearance and comparing the patient’s PET image to the model via a Z-score, which results in the generation of subject-specific abnormality maps summarising the pathology’s topographical distribution in the brain. Extending this work, I have proposed a strategy to validate the abnormality maps on several PET tracers and automatically detect dementia by using the abnormality maps as features to feed a linear support vector machine (SVM) classifier. The evaluation strategy has been carefully designed to guarantee unbiased results by using two nested cross-validation procedures to train the classifier and to optimise the hyperparameters. This strategy enables us to assess on a large dataset (~300 subjects selected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database) if the proposed subject-specific abnormality maps are able to extract for each individual the signal characteristic of abnormality from both FDG (marker of synaptic dysfunction) and Florbetapir (marker of amyloid deposition) PET data. The high classification accuracy obtained when using the abnormality maps as features (e.g. balanced accuracy of 91.6% when differentiating cognitively normal and Alzheimer’s disease subjects using FDG PET) demonstrates that the proposed pipeline is able to extract for each individual the signal characteristic of dementia from both FDG and Florbetapir PET data. Examples of abnormality maps and of a map showing the voxels the most relevant for the classification of cognitively normal and Alzheimer’s disease subjects are displayed in the figure below.
The classifier used in this work has been implemented as part of a broader effort of the Aramis Lab meant to develop a framework for transparent and reproducible evaluation of classification algorithms. In this context, I participate to the supervision of a PhD student, whose thesis objective is to develop and validate new machine learning approaches that can integrate data from multiple neuroimaging modalities (MRI and PET) in order to predict the evolution of patients from the earliest stages of Alzheimer’s disease. Preliminary results obtained with the classification framework that he developed have been presented at the MICCAI workshop on Machine Learning for Medical Imaging (paper, code) . The classification tools are available in Clinica, an open-source software platform for clinical neuroscience research studies developed by the Aramis Lab, to which I contribute, mainly by participating to the management of the project and to the writing of the documentation. I also co-supervise an engineer who develops software tools to process multi-modal medical images, with a specific focus on PET and MRI data.
- Burgos, N., Samper-González, J., Bertrand, A., Habert, M.-O., Ourselin, S., Durrleman, S., Cardoso, M.J., and Colliot, O.: Individual Analysis of Molecular Brain Imaging Data through Automatic Identification of Abnormality Patterns. In Molecular Imaging, Reconstruction and Analysis of Moving Body Organs, and Stroke Imaging and Treatment, LNCS, 10555: 13–22, Springer, 2017. doi:10.1007/978-3-319-67564-0_2