FreqDualNet: frequency-aware vision transformers for tumor segmentation
PeerJ Computer Science · 21.04.2026 · 12 · e3801
DOI: 10.7717/peerj-cs.3801 ↗Kurz gefasst
Fachartikel in der Zeitschrift PeerJ Computer Science, online erschienen am 21.04.2026. Verfasst von Jungro Lee und 1 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Research Foundation of Korea (NRF) through the Korean Government.
Abstract (Originalsprache)
Recent advances in deep learning models such as U-Net Transformer (UNETR) and Swin U-Net Transformer (SwinUNETR) have significantly improved tumor segmentation accuracy and local structural sensitivity in medical imaging. Notably, OrgUNETR model demonstrated that predicting both organs and tumors enhances tumor segmentation performance by leveraging their anatomical relationships. However, those models still struggle to capture the complex and irregular shapes of tumors that result from structural constraints or physiological pressure exerted by surrounding organs. These irregular morphologies often manifest as lobulated, or infiltrative margins, containing high frequency edge information that is challenging to represent using a convolutional network or transformer based architectures alone. To address this limitation, we propose FreqDualNet, a frequency aware extension of SwinUNETR and OrgUNETR, which explicitly integrates three types of frequency awareness modules into the network. FreqDualNet introduces three complementary frequency awareness modules: (i) Fourier based spectral feature transformation to enhance high frequency sensitivity, (ii) frequency guided feature fusion between organ and tumor representations, and (iii) high frequency residual refinement to sharpen tumor boundaries. By explicitly modeling frequency domain characteristics associated with tumor shape complexity, the proposed architecture improves sensitivity to fine-grained morphological variations. Experiments on the KiTS19 dataset demonstrate that FreqDualNet achieves a substantial improvement in tumor segmentation performance, increasing the tumor Dice score by 12.75% compared to SwinUNETR (0.5072 vs . 0.4499). On the Prostate158 dataset, FreqDualNet also demonstrated improved tumor segmentation performance, achieving a 6.70% increase in tumor Dice score compared to SwinUNETR (0.2282 vs . 0.2138). By directly utilizing the morphological differences between organs and tumors with frequency aware modules, FreqDualNet improves both segmentation accuracy and generalizability, with strong potential to assist tumor diagnosis, prognosis evaluation, and understanding tumor organ interactions. The code is available at: https://github.com/JungroLee/FreqDualNet .
Förderung
- National Research Foundation of Korea (NRF) through the Korean Government
Zitieren
APA
Lee, J., & Lee, M. (2026). FreqDualNet: frequency-aware vision transformers for tumor segmentation. PeerJ Computer Science, 12, e3801. https://doi.org/10.7717/peerj-cs.3801
Vancouver
Lee J, Lee M. FreqDualNet: frequency-aware vision transformers for tumor segmentation. PeerJ Computer Science. 2026;12:e3801. doi:10.7717/peerj-cs.3801
BibTeX
@article{Lee2026,
author = {Lee, Jungro and Lee, Minhyeok},
title = {FreqDualNet: frequency-aware vision transformers for tumor segmentation},
journal = {PeerJ Computer Science},
year = {2026},
volume = {12},
pages = {e3801},
doi = {10.7717/peerj-cs.3801}
}
EUVIMED