Connectome priors in deep neural networks to predict autism

We propose a novel element-wise layer for deep neural networks that incorporates general priors designed for connectomes. In contrast to regular images, connectomes, expressed as adjacency matrices, are composed of elements that capture a relationship between two brain regions. As each element in the connectome has an anatomical meaning that is consistent across samples, prior knowledge about the structure of the connectome can be encoded and used to regularize learning based approaches. Thus in this work, we introduce a novel trainable element-wise layer for deep neural networks, with data-dependent anatomically-informed prior regularization terms designed for connectomes. We validate our approach on 1013 functional connectomes from the autism brain imaging data exchange (ABIDE) dataset, and show that our proposed layer and regularization terms improves the accuracy of predicting patients with autism spectrum disorder from controls within a deep learning framework.