Скачать или смотреть OHBM 2022 | 168 | Talk | Bo Li | A high-resolution autoencoder for construction of interpretable b…

Title: A high-resolution autoencoder for construction of interpretable brain MRI endophenotypes.
Session: Talk
Speaker: Bo Li
In population neuroimaging studies it is especially important to select a right phenotype of interest for the analysis because of the high-dimensional nature of imaging data versus the relatively small number of observations (i.e., subjects). Selecting a suboptimal phenotype such as specific brain regions can lead to the loss of statistical power or misleading results. To eliminate the need and the bias of phenotype selection, unsupervised dimensionality reduction techniques such as principal component analysis (PCA) - are often applied to construct compact representations (endophenotypes) from the images. However, representations from PCA may not be easily interpretable. Also, preserving essential information from high-dimensional neuroimaging can be quite challenging and computationally intense. Here, we propose a first unsupervised autoencoder for interpretable dimensionality reduction of high-resolution neuroimaging data that preserves local and global brain structure. In two proof of principle experiments, we will demonstrate the method’s performance in decoding age and gender-related brain regions.