Feature-Level Fourier Amplitude Modulation to Reduce Domain-Specific Spectral Bias for Cross-Domain Polyp Segmentation
Journal of Imaging Informatics in Medicine · 12.11.2026
DOI: 10.1007/s10278-026-02480-3 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Imaging Informatics in Medicine, online erschienen am 12.11.2026. Verfasst von Haesung Oh und 2 weiteren Autor:innen. Beteiligte Einrichtung: Department of Artificial Intelligence, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea. Die Arbeit nennt 46 Literaturangaben. Gefördert von: Ministry of Science and ICT, South Korea.
Abstract (Originalsprache)
Accurate polyp segmentation is critical for early colorectal cancer detection, yet models trained on one endoscopic dataset often degrade during cross-domain evaluation. One major source of this gap is acquisition-induced appearance variation, where imaging settings and post-processing pipelines introduce color shifts that encourage models to rely on domain-specific cues. Existing image-level Fourier amplitude approaches reduce global appearance discrepancy but align the entire image spectrum before feature extraction, which risks coupling mucosal background statistics with polyp-relevant signals. We instead modulate the Fourier amplitude of encoder features. Specifically, we decompose the feature amplitude into low- and high-frequency bands and recalibrate their second-order amplitude statistics in a projected space, while preserving the original phase. Since spectral modulation can attenuate fine boundary-related high-frequency responses, we complement it with directional edge refinement guided by an auxiliary luminance signal. Extensive experiments across five polyp segmentation benchmarks under cross-domain protocols show consistent improvements over strong baselines. Quantitatively, the proposed method attains the highest Dice on all five benchmarks (0.930 on Kvasir-SEG and 0.938 on CVC-ClinicDB in-domain; 0.891, 0.805, and 0.830 on the unseen CVC-300, CVC-ColonDB, and ETIS datasets).
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Artificial Intelligence, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea.
- Department of Electrical, Electronic and Systems Engineering, Universiti Kebangsaan Malaysia, Bangi, 43600, Selangor, Malaysia.
- Department of Artificial Intelligence, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea. curseor@cau.ac.kr.
Förderung
- Ministry of Science and ICT, South Korea
Zitieren
APA
Oh, H., Zulkifley, M. A., & Lee, J. (2026). Feature-Level Fourier Amplitude Modulation to Reduce Domain-Specific Spectral Bias for Cross-Domain Polyp Segmentation. Journal of Imaging Informatics in Medicine. https://doi.org/10.1007/s10278-026-02480-3
Vancouver
Oh H, Zulkifley MA, Lee J. Feature-Level Fourier Amplitude Modulation to Reduce Domain-Specific Spectral Bias for Cross-Domain Polyp Segmentation. Journal of Imaging Informatics in Medicine. 2026. doi:10.1007/s10278-026-02480-3
BibTeX
@article{Oh2026,
author = {Oh, Haesung and Zulkifley, Mohd Asyraf and Lee, Jaesung},
title = {Feature-Level Fourier Amplitude Modulation to Reduce Domain-Specific Spectral Bias for Cross-Domain Polyp Segmentation},
journal = {Journal of Imaging Informatics in Medicine},
year = {2026},
doi = {10.1007/s10278-026-02480-3}
}
EUVIMED