Alzheimer's disease is the most common form of dementia and a major challenge for healthcare systems. Alongside clinical tests, magnetic resonance imaging shows promise to aid in the diagnostic process. Recent advances in computing power made processing of MR images using deep learning models feasible. The analysis of contemporary classification approaches showed that they mostly rely on structural MRI data acquired with MPRAGE sequences. In this thesis the applicability of additional image contrasts (FLAIR, R2, R2*, MTR) for Alzheimer’s disease classification with deep convolutional neural networks is considered and compared. Furthermore, the explanation method “deep Taylor decomposition” indicated that those networks might learn features introduced by the applied image preprocessing. Therefore, this thesis introduces a new method to mitigate these problems by using the generated heatmaps during training for regularization. The proposed method leads to identification of features in anatomically more plausible regions, while maintaining a similar classification accuracy compared to contemporary approaches.
|Qualification||Master of Science|
|Publication status||Published - 2020|
- Alzheimer’s disease
- deep learning
- relevance-guided training