MRI-based pattern recognition techniques in dementia diagnostics

• Applicant: Professor Dr. Stefan Klöppel
• Subject Area: Human Cognitive and Systems Neuroscience ; Cognitive, Systems and Behavioural Neurobiology ; Medical Physics, Biomedical Technology
• Term: from 2011 to 2017
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 196903235

Recent evaluations in Germany have shown that dementia diagnostics lack specificity. In a primary care setting, patients with dementia mostly receive a diagnosis of unspecific dementia which substantially limits their chances of receiving medication that was specifically developed for Alzheimer s disease (AD). Figures obtained from neurologists and psychiatrists outside specialized centers are similar. Although brain imaging is part of the routine workup of dementia diagnostics and could improve diagnostic accuracy, the interpretation of these images depends heavily on the experience of radiologists which may be limited outside specialized centers. In the initial funded phase, we demonstrated that automated image processing and pattern recognition methods applied to a MRI-scan of the brain can improve the diagnostic quality. The employed image processing and the support-vector machine (SVM) based classification algorithm were sufficiently robust to handle data from different scanner. On the other hand, we identified a substantial drop of diagnostic accuracy when applying a classifier trained on data from well-controlled multi-center studies to new data from our local memory clinic. To overcome this critical limitation, we plan to apply deep convolutional neuronal networks (CNN). We found these methods to outperform hand-crafted features in a range of applications. CNNs can be applied to the raw images without e.g. a segmentation into tissue types or non-linear registration to a reference image. Moreover, as they integrate all steps from raw data to diagnostic decision, these methods can use the misdiagnosing of training data to improve the extraction of relevant information. With the clinical application in mind, we also plan to develop a network that can combine multiple data types per subject but will not fail, when one type of data is not available from one subject (e.g. because visual impairment or anxiety led to an incomplete cognitive testing or imaging battery). Finally, we intend to assign an uncertainty to each diagnostic decision by the CNN to better inform clinicians about the reliability. All developed methods will be tested on existing data from the local memory clinic but will also be evaluated in a prospective study. To this end, the outcome of the CNN together with a visualization of the underlying data will be presented in the weekly case conferences of the clinic.

• DFG Programme: Research Grants