Abstract
Liver fibrosis, the progressive scarring that can advance to cirrhosis, is a common chronic liver condition whose severity is traditionally assessed by biopsy. MRI offers a non-invasive alternative through two complementary tasks: segmenting the liver and staging the severity of fibrosis. For such models to be clinically useful, they must generalise beyond the scanners they were trained on, since MRI appearance varies substantially across vendors. The 2026 CARE-Liver benchmark is built around this requirement: its 460 training cases come from three scanner vendors sharing no acquisition protocol, only 30 carry segmentation labels, and the test set originates from a fourth, unseen vendor. We present a unified study of both tracks in which every design decision is judged by leave-one-vendor-out (LOVO) evaluation rather than by in-distribution accuracy. For segmentation, we compare supervised and self-supervised (DINOv2) pretraining across state-of-the-art architectures, including ResEnc-L and the Primus-M transformer, and find that supervised pretraining followed by fine-tuning generalises best to the unseen vendor among the evaluated configurations. For staging, we evaluate a missing-modality transformer which is an architecture designed to fuse multiple MRI phases while remaining robust to phases that are absent at inference by substituting learned embeddings for missing input, and find that although it is competitive in-distribution, it collapses under LOVO, whereas transferring the pretrained segmentation encoders generalises substantially better. Our results indicate that, under vendor shift, external pretraining and encoder transfer matter more than architectural choice.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CARE_040.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=7BlGltKXWj
BibTex
@InProceedings{CenUmu_VendorRobust_MICCAISAT2026,
author = { Cengiz, Umut Alperen AND Uzunay, Emine Şevval Eş AND Erkol, Irmak AND Eckstein, Katharina AND Ulrich, Constantin AND Maier-Hein, Klaus H.},
title = { { Vendor-Robust Liver Segmentation and Fibrosis Staging from Multi-Phase MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}