Abstract
Low-field magnetic resonance imaging (MRI) can broaden access to pediatric neuroimaging, but automatic segmentation remains difficult because contrast is weak, signal-to-noise ratio is limited, and several target structures are small. For LISA 2026 Task~2, we developed a reproducible nnU-Net v2 pipeline for multi-structure segmentation of pediatric brain T2 MRI. Following prior LISA work, we retained the official left/right labels and evaluated residual-encoder models without mirroring. A controlled three-way ablation showed that naive mirroring reduced mean performance, whereas anatomy-aware mirroring with explicit left/right label exchange recovered most of the loss and did not differ clearly in Dice from no mirroring after paired multiplicity-adjusted testing. No coarse global left/right swaps were detected. Five-fold cross-validation on 79 training cases otherwise showed only minor differences among a plain nnU-Net, standard ResEnc-M, and a cubic-patch ResEnc-M. We therefore evaluated a heterogeneous ensemble that adds stronger augmentation, supervised external pretraining, and self-supervised initialization. A hippocampus-only model, inspired by the specialist strategy used in LISA 2025, was combined with the generalist ensemble through explicit label-level merge rules. The deployable hippo-first merge produced the numerically highest mean and hippocampal Dice scores among the deployable specialist variants, whereas the generalist-only ensemble remained better on the distance and volume metrics. After multiplicity adjustment the ensemble improvement was statistically significant, whereas the specialist merge produced no detectable change. The observed differences are small and metric-dependent. We consequently present the method as an incremental, reproducible challenge pipeline and do not claim that one configuration is uniformly superior.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_012.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=MqME7EvlG9
BibTex
@InProceedings{GagGia_Ensembled_MICCAISAT2026,
author = { Gagliardi, Gianni Manuel AND Graewert, Frederik AND Palm, Maximilian AND Deissler, Jonathan AND Ulrich, Constantin},
title = { { Ensembled Laterality-Aware Residual nnU-Net for Low-Field Pediatric Brain MRI Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17257},
month = {pending},
page = {pending}
}