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: | Mohammed Aburidi |
| Institution: | University of California Merced |
| Department: | Applied Mathematics |
| Country: | |
| Proposed Analysis: | Image recognition systems and deep learning algorithms have greatly impacted the field of medical diagnostics and image processing. Despite the large number of retrospective studies based on AI principles, only a few applications are now in clinical use. One main reason for that is the poor transferability of these models to data from different sources or acquisition protocols. This variation in acquisition protocols and data domains, results in different appearances of normal and diseased tissue in the images, even when the same preprocessing pipeline was used and is brought to a standard form. This makes these models not perform satisfactorily on data acquired with different protocols or in different domains. This, in turn, limits the use and thus transferability of models trained with large annotated datasets on a new dataset with a different domain or acquisition protocol. Therefore, the development of new methods and algorithms that are more generalizable to larger data sets from different sources and acquisition settings is crucial for clinical use. In this proposal, we propose a methodology to tackle this problem and aim to answer the following questions: 1. What are the necessary data preprocessing steps that help to make the model adaptable to a new domain? 2. What are the model components that are needed to be changed in order to achieve a higher level of generalizability? 3. How large should the data set be so the model achieves the desirable adaptation of the original model? We aim to build a novel task-Specific 3D deep learning model for domain adaptation, with necessary data preprocessing steps to maximize the generalizability of a model. As a first point, the model will be designed to solve an Alzheimer’s disease multi-class classification task that classifies a patient’s MRI image into one of the main four categories (Alzheimer’s disease, normal cognition, mild cognitive impairment, dementia due to other causes). This model will be trained using data from one domain and a similar acquisition protocol. Then, changes on the following levels should take place to address the above-stated three questions: first, changes on the data preprocessing level. For example, we plan to test with new normalization and pixel intensities standardization tools and see the effect on the model’s performance when medical imaging data from a different domain is tested. Secondly, to address the second question we will incorporate new modalities in the model architecture and study their effect on the performance when new data from different domains are used. Our hope is to come up with a reliable and robust 3D model trained on one modality and can be generalized to multiple modalities (domains). To this end, we reach out to the most influential project for the research of Alzheimer’s disease, ADNI, to give us access to MRI imaging datasets, which we will need to benchmark our model. Thank you very much and we are looking forward to hearing positively from you. |
| Additional Investigators |
