Abstract
Ultra-low-field (0.064 T) portable MRI is expanding pedi-
atric neuroimaging access in low-resource settings, but its low signal-to-
noise ratio and heavy artifact burden demand automated quality assur-
ance. We address Task 1a of the LISA 2026 challenge: per-image multi-
label ordinal grading (absent / mild / severe) of seven artifacts—noise,
zipper, positioning, banding, motion, contrast, distortion—on 0.064 T
T2-weighted volumes. Our method pairs a view-conditional 2.5D slab
representation, which respects the strong through-plane anisotropy, with
an ensemble of complementary ImageNet-pretrained backbones, each
trained with a different imbalance-aware or ordinal loss so that mem-
bers make decorrelated errors on the rare severe grades; an auxiliary
brain bounding-box head adds spatial regularization. Through a strict
nested-calibration analysis we show that per-artifact decision calibration
overfits, and that a wide, uncalibrated ensemble generalizes best. On
patient-level five-fold cross-validation we reach an overall weighted score
of 0.835; the same submission scores 0.831 weighted-mean and 0.841 ac-
curacy on the hidden validation set. A broad ablation shows only pretraining and ensemble diversity transfer
reliably; the tight leaderboard clustering indicates a label-noise ceiling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_008.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=z45IJvlWox
BibTex
@InProceedings{KolGov_DiverseLoss_MICCAISAT2026,
author = { Kolli, Govinda AND Dukre, Adinath Madhavrao AND Razzak, Imran},
title = { { Diverse-Loss Backbone Ensembling for Multi-Label Artifact Quality Assessment of Ultra-Low-Field Pediatric Brain MRI } },
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}
}