We divide the dataset into two atlases so that the features are from the same domain. We build a correspondence from one atlas to another, searching for the most similar samples in different atlases. The similarity of the samples is determined by the degree of correlation of the features.
Having obtained an unambiguous correspondence between samples in different atlases, we cluster them in the Brainnetome atlas, discarding features with an index of 210 and higher, since it was worse with them. We build Connectivity Matrixes (hereinafter CM) from the Nilearn library for samples using the kind="precision" parameter and using connectivity from the other hemisphere for features from one hemisphere (so clustering is performed unambiguously).
Having received CM, we pull them into a vector and use them to determine the similarity of samples among themselves, counting the Pearson correlation between the obtained vectors. Next, we build a graph with weights for clustering.
- The data was augmented: the time line of the images was divided into sub-sections of length 80 in increments of 40
- We did undersampling: according to the Beijing dataset, in order to withstand 50/50 class balancing everywhere
- Training: a LogReg with L1 regularization is pre-trained, after which unimportant features are cut off (where weights <= 1e-5); then we reduce the dimension using PCA and train another LogReg on the final data set
- Arkhipov D.A.
- Mirantsov E.S.
- Kochyan L.V.