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: | Marek Kraft |
| Institution: | Poznan University of Technology |
| Department: | Institute of Robotics and Machine Intelligence |
| Country: | |
| Proposed Analysis: | In terms of metrics, the best-performing methods for finding specific regions in images or volumes (segmentation) are currently based on deep neural networks. The neural networks are essentially a supervised machine learning algorithm. Before they can perform their intended function, they need to be trained using data that contains a sufficient number of templates. In the case of WMH segmentation, the templates have a form of manual annotations overlaid on top of the original scan. Hand annotation is a time-consuming process, as accurate delineation of the areas of interest requires significant effort, repeatability and consistency. Moreover, unlike annotation of everyday objects and scenes it requires some degree of medical expertise. As a result, publicly available data with accurate annotations ow WMHs is really scarce. The largest publicly available dataset is the WMH segmentation challenge dataset (https://wmh.isi.uu.nl/), containing 60 annotated scans. With neural network-based methods considered, this is a very small dataset. This turns the attention towards methods and tools that can deal with this data deficiency. Currently, we are in the process of developing pretext tasks to generate pseudo-labels and the results are encouraging. Below is an example of a raw slice of FLAIR sequence MRI scan and sample masks generated by pretext task using two different methods – clustering and SLIC superpixel grouping using expectation maximization. We hope that our methods will enable dealing with data deficiency without resorting to laborious manual annotations. Our plan is to thoroughly and methodically evaluate the self-supervised training approaches. In the long term, we also plan to use ADNI for white matter lesion quantification and longitudinal studies. ADNI 3 data with high resolution and good imaging quality is a very interesting source of data in this context. |
| Additional Investigators |
