Path-Weights and Layer-Wise Relevance Propagation for Explainability of ANNs with fMRI Data

The application of artificial neural networks (ANNs) to functional magnetic resonance imaging (fMRI) data has recently gained renewed attention for signal analysis, modeling the underlying processes, and knowledge extraction. Although adequately trained ANNs characterize by high predictive performance, the intrinsic models tend to be inscrutable due to their complex architectures. Still, explainable artificial intelligence (xAI) looks to find methods that can help to delve into ANNs' structures and reveal which inputs most contribute to correct predictions and how the networks unroll calculations until the final decision. Several methods have been proposed to explain the black-box ANNs' decisions, with layer-wise relevance propagation (LRP) being the current state-of-the-art. This study aims to investigate the consistency between LRP-based and path-weight-based analysis and how the network's pruning and retraining processes affect each method in the context of fMRI data analysis. The procedure is tested with fMRI data obtained in a motor paradigm. Both methods were applied to a fully connected ANN, and to pruned and retrained versions. The results show that both methods agree on the most relevant inputs for each stimulus. The pruning process did not lead to major disagreements. Retraining affected both methods similarly, exacerbating the changes initially observed in the pruning process. Notably, the inputs retained for the ultimate ANN are in accordance with the established neuroscientific literature concerning motor action in the brain, validating the procedure and explaining methods. Therefore, both methods can yield valuable insights for understanding the original fMRI data and extracting knowledge.