Anomaly Detection for Structural and Functional Connectivity in Glioma Patients

Abstract

Brain connectivity, quantified with diffusion MRI (structural connectivity, SC) and resting-state functional MRI (functional connectivity, FC), can offer crucial insights into glioma-brain network interactions. Currently, no standardized approach exists to integrate information from FC and SC and to identify potential tumor-induced abnormalities at the single-patient level. Variational autoencoders (VAEs) have been shown to be promising for learning the distribution of features representing a healthy brain and deviations thereof and can naturally be applicable to multiple modalities. This study explores the potential of VAE to integrate FC and SC and detect multimodal anomalies in brain connectivity in glioma patients. The VAE is trained on concatenated FC-SC healthy data to learn how to reconstruct normative connectivity patterns. After ad hoc transfer learning, the model parameters are applied to the oncological dataset, to obtain the healthy version of the pathological matrices. Given the healthy, pathological, and reconstructed matrices, a statistic is developed with the goal of identifying specific alterations in SC, FC, and their FC + SC integration in glioma patients. SC, FC, and FC + SC abnormalities are compared with each other to explore their interplay and their link with tumor and surrounding brain tissues. Results show that FC is more sensitive to alterations distant from the tumor, while SC is more affected in its vicinity. Then, the alterations identified by FC are generally more in agreement with the alterations identified by FC + SC compared with those highlighted by SC. Moreover, SC abnormalities never overlap with FC + SC out of the tumor, and FC and SC single impairments partially overlap within the tumor core and never overlie in other brain tissues. This information could facilitate patient stratification, prognostic modeling, and personalized treatment planning.