Identifying Neural Signatures of Learning in Functional Ultrasound via Tensor Decomposition

Abstract

Understanding the mechanisms that underpin learning effects remains a central goal in neuroscience, with functional neuroimaging offering a powerful avenue for observing brain activity. This work is situated within the domain of haemodynamic functional neuroimaging, with a specific focus on functional ultrasound (fUS)—a relatively novel modality that combines high spatio-temporal resolution and portability. The goal of this project is to investigate the capacity of tensor-decomposition-based techniques to extract meaningful, interpretable representations of stimulus-evoked learning effects from a multi-subject fUS dataset while being robust to the non-idealities present in said data. While numerous approaches exist for analysing neuroimaging data, many are limited by scalability, interpretability, or an inability to capture changing neurological patterns. This work is thus motivated by the need for methods capable of handling large, dynamic datasets while extracting interpretable results. Tensor decompositions offer such a framework but their ability to capture subject-specific time-varying effects and their application to functional ultrasound is still under-explored. Several tensor decomposition algorithms are examined, assessed in terms of their learning effect extraction capabilities and robustness to non-idealities, and a novel shifted canonical polyadic decomposition variant is developed to address time-varying effects. Results from synthetic data analysis demonstrate that tensor decompositions are robust to non-idealities in functional ultrasound data (to varying degrees) and can recover interpretable latent components. Due to the data's partial scan coverage, no novel signatures of learning were identified in the real data; however, the extracted components further substantiate the potential of these algorithms and highlight a promising direction for analysing learning effects in functional neuroimaging data.