Abstract
For nucleus segmentation pipelines built without large annotated datasets or high-memory GPU infrastructure, a simple VGG U-Net encoder matches purpose-built transformer alternatives at a fraction of Swin-T’s parameter count. We present a controlled comparison of VGGstyle, ResNet-style, and Swin Transformer encoders within a shared UNet framework, evaluated on PanNuke (∼5K training patches, 19 tissue types) and MoNuSeg (37 training whole-slide images). On PanNuke, VGG (Dice 0.854) matches ResNet (0.853) to within 0.002 Dice and outperforms pretrained Swin by 3.0 Dice points (0.824). On MoNuSeg, VGG (0.794) reaches published MedT-level performance—a transformer designed specifically for small medical datasets—while outperforming pretrained Swin by 3.6 Dice points (0.758). VGG is also the fastest and most memory-efficient of the learned models we test, making it Paretodominant over Swin-T on accuracy and inference cost. Lightweight fromscratch architectures remain the practical choice for edge microscopy, veterinary and rare-disease research, and non-H&E protocols. Of the three encoders we compare, a simple VGG is the strongest.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMERGE_018.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=ZiRs1areDR
BibTex
@InProceedings{RolDan_When_MICCAISAT2026,
author = { Rollo, Danny},
title = { { When Simple Wins: Lightweight CNN Encoders for Resource-Constrained Nucleus Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17266},
month = {pending},
page = {pending}
}
