Abstract
Deep learning segmentation models have demonstrated strong performance across medical imaging tasks, yet their practical deployment in resource-constrained clinical environments — particularly in low- and middle-income countries — remains limited by high computational demands, large model sizes, and GPU dependence. We present a training-free method for dataset-adaptive architectural compression of the nnU-Net v2 ResEncM pipeline, implemented by applying the Jacobian sensitivity analysis of the XTinyU-Net framework within the nnU-Net v2 pipeline at initialisation time. By identifying the onset of representational saturation across both channel width and residual block depth configurations, the method selects the smallest stable architecture for each dataset without any additional training. The approach was validated across three datasets of increasing clinical complexity: the ACDC cardiac segmentation dataset, the BraTS-Africa glioma dataset, and the BraTS-METS brain metastasis dataset. Across all three, the lightweight models reduced model size by 94.9–96.9% (796.9 MB to 25.1–40.7 MB) and CPU inference time by 82.7–90.1%, reducing BraTS-METS inference from over 11 minutes to under 2 minutes on CPU-only hardware. Segmentation accuracy was preserved or improved: the lightweight model matched or marginally exceeded the baseline on BraTS-Africa, outperformed it on BraTS-METS across all sub-regions (ET DSC: 0.534→0.602, NETC: 0.547→0.624, SNFH: 0.497→0.561), and produced fewer gross boundary failures on ACDC despite marginally lower mean DSC. These results demonstrate that training-free sensitivity-guided compression is a principled and dataset-adaptive pathway to clinically deployable segmentation, with no trade-off between performance and deployability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_012_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=9JDuJfc823
BibTex
@InProceedings{BooWil_Towards_MICCAISAT2026,
author = { Boonzaier, Willem P. E. AND Pretorius, Izak S. AND Musah, Toufiq AND Strauss, Lourens J. AND Anazodo, Udunna C.},
title = { { Towards CPU-Deployable nnU-Net: Training-Free Sensitivity-Guided Compression Across Clinical Domains } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}