BCIwiki (bciwiki.com) — According to preprint information supplied by arXiv, the DenoFlow study was uploaded to arXiv on 2026-09-24 and has not been peer reviewed. Authors Zhentao He, Ziwei Wang, and Dongrui Wu propose using flow matching to remove physiological artifacts from steady-state visual evoked potential (SSVEP) signals, with supervision that also accounts for decoding performance after denoising.
DenoFlow follows a rectified-flow formulation: a field network regresses the velocity along a straight path between contaminated and clean trials, and denoising integrates that field forward from the observation. The encoder-decoder network receives the contaminated trial at every layer and the path position at its bottleneck. A jointly trained classifier supervises the integrated output. The authors explain that using the observation as both the conditioning input and the integration starting point avoids generating trials from noise and removes the need for an adversarial min-max training game.
To construct paired data for datasets without clean ground truth, the researchers injected physiological artifacts from recorded electromyography (EMG) and electrooculography (EOG) signals under controlled signal-to-noise targets. The authors report that, across two public SSVEP datasets evaluated with five popular SSVEP decoders, DenoFlow outperformed seven baseline denoising models in both signal fidelity and downstream decoding accuracy.