Developing Replicable model for Real-World Functional Neuroimage

1394 

Abstract Submission 

Jiaqi Ding¹, Tingting Dan¹, Ziquan Wei¹, Guorong Wu¹

¹University of North Carolina at Chapel Hill, Chapel Hill, NC

Jiaqi Ding  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Tingting Dan  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Ziquan Wei  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Guorong Wu  
University of North Carolina at Chapel Hill
Chapel Hill, NC

Neuroimaging techniques have revolutionized our capacity to understand the neurobiological underpinnings of behaviorin-vivo[1]. Leveraging an unprecedented wealth of public neuroimaging data, there is a surging interest to answer novel neuroscience questions using machine learning techniques. Despite the remarkable successes in existing deep models, current state-of-the-arts[2] have not yet recognized the potential issues of experimental replicability arising from ubiquitous cognitive state changes, which might lead to spurious conclusions and impede generalizability across neuroscience studies. In this work, we first dissect the critical (but often missed) challenge of ensuring prediction replicability in spite of task-irrelevant functional fluctuations. Then, we formulate the solution as a domain adaptation where we devise a cross-attention mechanism with discrepancy loss in a Transformer backbone.

In our domain cross-attention Transformer[8], as shown in Figure 1, first, we introduce the cross-attention layer, an alternating module comprising self-attention in the source domain and cross-attention based on source and target domain. The integration of cross-attention (shown in the top right corner of Figure 1) facilitates the model's ability to relate source domain information to the target domain. Specifically, suppose the output of target domain's previous-layer constitutes the query Q_T, while the outputs from the source domain form the key K_S and value V_S, so that the model achieves enhanced information exchange and contextual representation between the two domains and minimizes the inter-domain dependencies. Then, we propose an adjusted discrepancy loss that can align the predicted probability vectors from the source and target domains, ensuring that they exhibit similar patterns in a shared feature space. We first propose a novel mathematical approach to update the correspondence between the predicted vectors of the two domains within the same batch. Then, we compute the mean discrepancy loss between the predicted probabilities in the source domain and the updated probabilities in the target domain.

We have evaluated the cognitive task recognition accuracy and consistency on both test and retest functional neuroimages from the Human Connectome Project. Our working memory dataset is part of the HCP-Task dataset[7], it provides valuable insights into the brain's activity during working memory tasks, including two distinct scanning sessions referred to as test data (scan 1) and retest data (scan 2). In experiments, our primary objective was to delve into the model replicability. To this end, we employed the bidirectional validation framework. Initially, we train the model on the test data and test on the retest data (i.e., test-retest experiment). This process was then reversed (i.e., retest-test experiment). By assessing the divergences in the outcome distribution patterns between these distinct scans, we aimed to unravel the model's capacity to equivalently discern the inherent features of both datasets, enabling us to make sure whether it indeed grasped the task-specific features in the disparate scans. Figure 2 (last column, whole brain) shows the task recognition accuracy (for a total of eight tasks) by different methods[3][4][5][6]. It is clear that our method outperforms all the comparison methods (at a significant level p<0.0001) in bidirectional validation situations.

We bring attention to a critical issue of model replicability in consistently linking dynamic neural activity with cognition and behaviors. Upon identifying the prominent obstacle of limited generalizability within existing deep models used in fMRI research, we present a practical solution to address the replicability issue using a novel cross-attention transformer for test-retest scans of multi-session fMRI data, indicating great applicability of our data-driven approach in various neuroscience studies.

Working Memory

Classification and Predictive Modeling ¹

Methods Development ²

BOLD fMRI

Computational Neuroscience

FUNCTIONAL MRI

Machine Learning