Model combining topological descriptors with patch based MR imaging features.
This is a work in progress repositroy by F. Hensel and S. Brueningk according to the initial description in https://arxiv.org/abs/2011.06531. This code will be further imporved and hence may sightly deviate from the original description.
In out analysis we used T1-weighted MR images for AD and CN subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Data was preprocessed as described in the archive article using the fmriprep pipeline. For the creation of persistence images, we first calculated the persistence diagrams of the full MRIs using dipha and then subsequently computed the persistence images using persim.
The function run.py contains the code to run the image-patch-based 3D-CNN, the TDA 2D-CNN, and a combined model using both topoligical descriptors and a 3D image patch. The information for all 216 patched can be combined in a logistic regression model (ensemble model 1), whereas the preclassification layer encodings of the TDA 2D-CNN and single patch 3D-CNN can used as features for a single dense layer (ensemble model 2).