Template-based graph registration network for boosting the diagnosis of brain connectivity disorders.

Brain graphs are powerful representations to explore the biological roadmaps of the human brain in its healthy and disordered states. Recently, a few graph neural networks (GNNs) have been designed for brain connectivity synthesis and diagnosis. However, such non-Euclidean deep learning architectures might fail to capture the neural interactions between different brain regions as they are trained without guidance from any prior biological template-i.e., template-free learning. Here we assume that using a population-driven brain connectional template (CBT) that captures well the connectivity patterns fingerprinting a given brain state (e.g., healthy) can better guide the GNN training in its downstream learning task such as classification or regression. To this aim we design a plug-in graph registration network (GRN) that can be coupled with any conventional graph neural network (GNN) so as to boost its learning accuracy and generalizability to unseen samples. Our GRN is a graph generative adversarial network (gGAN), which registers brain graphs to a prior CBT. Next, the registered brain graphs are used to train typical GNN models. Our GRN can be integrated into any GNN working in an end-to-end fashion to boost its prediction accuracy. Our experiments showed that GRN remarkably boosted the prediction accuracy of four conventional GNN models across four neurological datasets.

Multigraph classification using learnable integration network with application to gender fingerprinting.

Multigraphs with heterogeneous views present one of the most challenging obstacles to classification tasks due to their complexity. Several works based on feature selection have been recently proposed to disentangle the problem of multigraph heterogeneity. However, such techniques have major drawbacks. First, the bulk of such works lies in the vectorization and the flattening operations, failing to preserve and exploit the rich topological properties of the multigraph. Second, they learn the classification process in a dichotomized manner where the cascaded learning steps are pieced in together independently. Hence, such architectures are inherently agnostic to the cumulative estimation error from step to step. To overcome these drawbacks, we introduce MICNet (multigraph integration and classifier network), the first end-to-end graph neural network based model for multigraph classification. First, we learn a single-view graph representation of a heterogeneous multigraph using a GNN based integration model. The integration process in our model helps tease apart the heterogeneity across the different views of the multigraph by generating a subject-specific graph template while preserving its geometrical and topological properties conserving the node-wise information while reducing the size of the graph (i.e., number of views). Second, we classify each integrated template using a geometric deep learning block which enables us to grasp the salient graph features. We train, in end-to-end fashion, these two blocks using a single objective function to optimize the classification performance. We evaluate our MICNet in gender classification using brain multigraphs derived from different cortical measures. We demonstrate that our MICNet significantly outperformed its variants thereby showing its great potential in multigraph classification.

Quantifying the reproducibility of graph neural networks using multigraph data representation.

Graph neural networks (GNNs) have witnessed an unprecedented proliferation in tackling several problems in computer vision, computer-aided diagnosis and related fields. While prior studies have focused on boosting the model accuracy, quantifying the reproducibility of the most discriminative features identified by GNNs is still an intact problem that yields concerns about their reliability in clinical applications in particular. Specifically, the reproducibility of biological markers across clinical datasets and distribution shifts across classes (e.g., healthy and disordered brains) is of paramount importance in revealing the underpinning mechanisms of diseases as well as propelling the development of personalized treatment. Motivated by these issues, we propose, for the first time, reproducibility-based GNN selection (RG-Select), a framework for GNN reproducibility assessment via the quantification of the most discriminative features (i.e., biomarkers) shared between different models. To ascertain the soundness of our framework, the reproducibility assessment embraces variations of different factors such as training strategies and data perturbations. Despite these challenges, our framework successfully yielded replicable conclusions across different training strategies and various clinical datasets. Our findings could thus pave the way for the development of biomarker trustworthiness and reliability assessment methods for computer-aided diagnosis and prognosis tasks. RG-Select code is available on GitHub at https://github.com/basiralab/RG-Select.

Predicting cognitive scores with graph neural networks through sample selection learning.

Analyzing the relation between intelligence and neural activity is of the utmost importance in understanding the working principles of the human brain in health and disease. In existing literature, functional brain connectomes have been used successfully to predict cognitive measures such as intelligence quotient (IQ) scores in both healthy and disordered cohorts using machine learning models. However, existing methods resort to flattening the brain connectome (i.e., graph) through vectorization which overlooks its topological properties. To address this limitation and inspired from the emerging graph neural networks (GNNs), we design a novel regression GNN model (namely RegGNN) for predicting IQ scores from brain connectivity. On top of that, we introduce a novel, fully modular sample selection method to select the best samples to learn from for our target prediction task. However, since such deep learning architectures are computationally expensive to train, we further propose a learning-based sample selection method that learns how to choose the training samples with the highest expected predictive power on unseen samples. For this, we capitalize on the fact that connectomes (i.e., their adjacency matrices) lie in the symmetric positive definite (SPD) matrix cone. Our results on full-scale and verbal IQ prediction outperforms comparison methods in autism spectrum disorder cohorts and achieves a competitive performance for neurotypical subjects using 3-fold cross-validation. Furthermore, we show that our sample selection approach generalizes to other learning-based methods, which shows its usefulness beyond our GNN architecture.