Abstract
Accurate segmentation of colon glands and retinal vessels requires capturing fine local texture and long-range spatial context. Convolutional U-Net variants aggregate global information only indirectly, while window-based attention remains locally constrained within each layer. Fourier token mixers offer global spatial support at $\mathcal{O}(N\log N)$ cost, but frequency-indexed spectral kernels scale in parameter count with the number of retained modes. We propose the Neural Field Fourier Token Mixer (NFFTM), which generates a Fourier kernel on-the-fly using an implicit neural representation conditioned on continuous frequency coordinates. This decouples parameter count from spectral resolution, enabling the INR architecture to serve as a tunable spectral prior. Integrated into a SwinUNETR-style hierarchical encoder–decoder, NFFTM is evaluated on GlaS and FIVES using SIREN, WIRE, FINER, and FF-MLP INRs. It achieves competitive performance against convolutional and windowed-attention baselines while using the lowest measured FLOPs/pixel in our controlled evaluation. Source code is available at: \url{https://github.com/nielsvyncke/nf-ftm-offgrid}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Off_Grid_036.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=SUuSEXtZ12
BibTex
@InProceedings{VynNie_Neural_MICCAISAT2026,
author = { Vyncke, Niels AND Ashtari, Pooya AND Pizurica, Aleksandra},
title = { { Neural Field Fourier Token Mixers for Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}
