DeepCor provides a flexible, participant-level denoising solution for fMRI that does not require large datasets or task information. It improves signal quality across simulations and real data, supporting clearer, more reliable interpretation of neural activity. 7/7

01.12.2025 14:31 — 👍 0    🔁 0    💬 0    📌 0

In category-localizer data, DeepCor enhances correlations between the BOLD signal and face stimuli in fusiform face area (FFA) 215% better than CompCor. 6/7

01.12.2025 14:31 — 👍 0    🔁 0    💬 0    📌 0

Using BrainIAK’s physiologically informed simulator, DeepCor achieves markedly higher R-squared values than CompCor, up to 339% improvement, under noise conditions that mimic aspect of real fMRI acquisition. 5/7

01.12.2025 14:30 — 👍 0    🔁 0    💬 0    📌 0

In simple synthetic data, DeepCor consistently recovers the ground truth better than CompCor, even when noise is nonlinear or very strong, where CompCor struggles to remove sinusoidal noise. 4/7

01.12.2025 14:30 — 👍 0    🔁 0    💬 0    📌 0

We extract time series from ROIs (signal+noise) and RONIs (noise-only), train a contrastive VAE to isolate noise features, and reconstruct denoised signals by zeroing out noise latents. 3/7

01.12.2025 14:30 — 👍 0    🔁 0    💬 0    📌 0

Grateful for the chance to work with Aidas Aglinskas and @steanze.bsky.social on this project. 2/7

01.12.2025 14:29 — 👍 0    🔁 0    💬 0    📌 0

DeepCor: denoising fMRI data with contrastive autoencoders - Nature Methods DeepCor is a deep-learning-based denoising approach for task-based and resting-state fMRI data that can be used even for single participants.

Excited to share our new Nature Methods paper on DeepCor, a contrastive autoencoder for denoising fMRI. DeepCor separates neural signal from noise at the voxel level, works for single participants, and outperforms CompCor by up to 215% on real data. 1/7

doi.org/10.1038/s415...

01.12.2025 14:28 — 👍 2    🔁 0    💬 6    📌 1