A Multi-Domain Noise-Aware 2.5-D UNet++ Framework With FiLM Conditioning and Spectral Gating for MRI Denoising
IEEE Access · 2026 · 14 · 114743-114774
DOI: 10.1109/access.2026.3717058 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von B. Uthra und 1 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Article Processing Charge, Vellore Institute of Technology.
Abstract (Originalsprache)
Denoising of Magnetic Resonance Imaging (MRI) remains challenging due to the presence of heterogeneous non-Gaussian noise introduced during image acquisition, reconstruction, and varying imaging conditions. Existing deep learning-based denoising methods often rely on simplified noise assumptions and may struggle to generalize across diverse noise domains. To address this limitation, a 2.5D noise-aware MRI denoising framework is proposed that combines realistic noise synthesis with adaptive feature learning. MRI studies annotated as ‘No Findings’ in the ACNS0332 metadata were used as lesion-free reference images for realistic noise synthesis. The proposed FiLM-conditioned UNet++ architecture incorporates spectral gating to jointly model spatial and frequency-domain information. In addition, residual learning, physics-aware constraints, and adaptive loss gating are employed to improve training stability and reconstruction quality. Experimental evaluation on an unseen test set containing multiple noise regimes demonstrates effective denoising performance, achieving a mean PSNR of $32.40~\pm ~5.80$ dB across diverse noise conditions, while maintaining low perceptual error and preserving anatomical information. The framework also achieves a real-time inference speed of 35.50 FPS. These results demonstrate the potential of combining multi-domain noise modeling, adaptive conditioning, and spectral learning for robust MRI denoising.
Förderung
- Article Processing Charge
- Vellore Institute of Technology
Zitieren
APA
Uthra, B., & Sivashankari, R. (2026). A Multi-Domain Noise-Aware 2.5-D UNet++ Framework With FiLM Conditioning and Spectral Gating for MRI Denoising. IEEE Access, 14, 114743-114774. https://doi.org/10.1109/access.2026.3717058
Vancouver
Uthra B, Sivashankari R. A Multi-Domain Noise-Aware 2.5-D UNet++ Framework With FiLM Conditioning and Spectral Gating for MRI Denoising. IEEE Access. 2026;14:114743-114774. doi:10.1109/access.2026.3717058
BibTeX
@article{Uthra2026,
author = {Uthra, B. and Sivashankari, R.},
title = {A Multi-Domain Noise-Aware 2.5-D UNet++ Framework With FiLM Conditioning and Spectral Gating for MRI Denoising},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {114743--114774},
doi = {10.1109/access.2026.3717058}
}
EUVIMED