A Multi-Echo-Informed Multimodal Spatio-Temporal CNN for ICA Component Classification in fMRI Denoising

Abstract

Multi-echo functional MRI (fMRI) improves signal quality by exploiting the fact that true blood-oxygenation-level-dependent (BOLD) signals decay with T2* across echo times while many artifacts do not. By decomposing multi-echo fMRI data using independent component analysis (ICA), each component’s signal decay across echo times can be used to classify it as BOLD-like or non-BOLD-like. The tedana software classifies ICA components using metrics that capture this differential decay via rule-based decision trees. However, approximately 66% of components in our cohort were not classified with high confidence using tedana, leaving residual artifacts that limit downstream analysis sensitivity. To address this gap, we developed a multimodal convolutional neural network (CNN) that jointly learns from the spatial maps and time series of ICA components of optimally combined resting-state multi-echo data, using training labels derived from these multi-echo decay metrics, to more accurately classify these components. Trained on 12,750 high-confidence components from 229 subjects with perfect class balance, the model achieved 98.6% accuracy on held-out data and flagged as artifacts 10,704 components (33.3% of the non-high-confidence set) that tedana had accepted. Critically, denoising with CNN classifications significantly improved network-based contrast-to-noise ratio (CNR) by 6.20% (p=0.011), with six of seven resting-state networks showing gains. Furthermore, CNN-based denoising increased network modularity (Louvain Q +6.8%, p=0.0006) and showed numerically higher CNR test-retest reliability (ICC: 0.664 vs. 0.622). Ablation studies showed that the CNN’s CNR advantage arises from retaining components that tedana rejects, while the additional components it rejects can be removed without loss of network-level signal. On two independent external datasets, the CNN generalized without retraining: CNR gains were directionally consistent but non-significant on a resting-state cohort (+2.6%, p=0.36), whereas on a more dissimilar task-fMRI cohort (four echoes, different scanner, TR = 1.615 s) it significantly improved network CNR over tedana (+16.7%, p<0.001) while performing comparably to tedana in tSNR and task activation, which was enhanced over the non-denoised baseline. SmoothGrad analysis showed broad cortical saliency consistent with resting-state networks, strongest at inferior boundaries, for accepted components, and lower-magnitude edge saliency for rejected ones, confirming neurobiologically meaningful feature learning. These findings demonstrate that learned spatiotemporal features provide superior multi-echo ICA classification compared to traditional metric-based decision trees, offering an interpretable, data-driven solution for multi-echo fMRI denoising.