Abstract
Brain metastases’ small, heterogeneous nature makes accu-
rate segmentation challenging. Volumetric metrics conceal critical fail-
ures: on BraTS-METS 2026, 3D U-Net smooth out small lesions, achieves
50.49% voxel Dice but only 4.89% lesion-wise Dice. Conversely, 2D mod-
els forfeit inter-slice consistency (2.28% lesion-wise Dice). The binding
constraint is inductive bias, not capacity. Given sparse targets and lim-
ited data without natural-image pre-training, feature transformations
must remain conservative with local spatial priors. We propose ACORN,
a lightweight 2.5D parallel multi-resolution network. Processing three
adjacent slices per MRI modality, it captures inter-slice context without
full 3D overhead. Parallel branches and cross-scale fusion preserve high-
resolution details and global context, while a lightweight decoder mini-
mizes computational cost. Three modules enhance small-lesion represen-
tations: Prog-AsymConv progressively activates zero-initialized asym-
metric branches, re-parameterizing them for inference; DG-CondDW
dynamically fuses expert kernels before depthwise convolution; and RAV-
SS refines selective-scan features via a parameter-efficient residual adapter
with a learnable scaling factor. With 2.23M parameters (5% of nnU-Net)
and 22.1 GFLOPs (1.9% of nnU-Net), ACORN achieves leading voxel
Dice (53.27%) and ASSD (10.53). Crucially, it attains 36.88% lesion-
wise Dice — 5.1×the strongest non-nnU-Net baseline, and exceeds nnU-
Net on the resection-cavity sub-region.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_030.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{LinZhi_ACORN_MICCAISAT2026,
author = { Lin, Zhi-Xun AND Wei, Chen-An AND Ting, Yu-Hsin AND Peng, Alina AND Chen, Chien-Chang},
title = { { ACORN: Asymmetric Conditional Overlapping RAVSS Network for Segmenting Small Brain Metastases } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17253},
month = {pending},
page = {pending}
}