There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | Ben Philps |
| Institution: | University of Edinburgh |
| Department: | School of Informatics |
| Country: | |
| Proposed Analysis: | In this project, machine learning tools will be used to evaluate and anlayse small vessel disease biomarkers (SVD). Neuroimaging has played a vital role in determining common markers of small vessel disease (SVD), and computer vision techniques are key to optimising diagnosis and intervention for patients. However, the automated assessment (i.e. segmentation and characterisation) of the neuroradiological features of SVD using machine learning tools such as Convolutional neural networks are hampered by reliance on acquisition parameters and scanner characteristics, available imaging modalities and non-reliable or missing ground truth. This represents a test-time shift in the input distribution used for Ml models, and presents the problem of Domain Generalization (DG). Furthermore, current methods are limited to assessing the current disease state, as opposed to disease progression. he adaption of AI tools in clinical practice is hampered by perormance degredation as input distributions shift away from the training distribution. This project will develop and explore the utility of DG, robustness and uncertainty quantification (UQ) techniques for bridging the gap between training and clinical test time performance for a variety of important SVD imaging analysis tasks (including assessment of white matter hyperintensities, stroke lesions, lacunes, perivascular spaces and microhaemorrhages). This requires broad, heterogeneous clinical data to identify inputs that risk model error or fairness issues. Furthermore, we will explore the utility of UQ techniques for improving the downstream prediction of disease characterization (e.g WMH fazekas metrics). A UQ technique for segmentation demonstrates the range of possible segmentations for a particular SVD feature type. This is particularly useful for WMH segmentations where deep isolated WMH are poorly localized or images with a low disease burden where a deterministic model may overestimate the lesion burden). Finally, by leveraging longitudinal data, unsupervised machine learning methods will be used to develop a model of brain ageging. We will then use the latent representations induced by this model, combined with outputs of our existing segmentation models and UQ feature maps, to develop a model of disease progression by predicting lesion segmentation maps for brain scans at later timepoints. These tasks form an important part of developing an automated personalised neuroradiological report to assist clinical practice. |
| Additional Investigators |
