Abstract
Ultra-low-field (ULF) MRI enables portable, low-cost, point of-
care neuroimaging, but its low signal-to-noise ratio, reduced tissue
contrast, anisotropic resolution, and acquisition artifacts hinder automated
analysis of pediatric brain scans. We present a uni ed framework
for Task 1 of the LISA 2026 Challenge, covering both automated artifact
assessment (Task 1a) and image quality improvement (Task 1b).
For Task 1a we formulate multi-artifact grading as an ordinal problem:
a 2.5D ConvNeXtGRU model predicts three-level severity for seven artifacts
through a cumulative-link head, with a subject-wise 5-fold ensemble.
For Task 1b, where paired clean targets are unavailable, we
train a metadata-preserving residual 3D U-Net on online-synthesized
degradedclean pairs; it predicts a bounded, identity-initialized correction
regularized by the frozen Task 1a model as an in-domain critic restricted
to intensity-domain artifacts, and restored volumes are written
back into the original NIfTI grid unchanged. On held-out proxy
validation the restorer reduces FID from 106.8 to 41.6, pooled LPIPS
from 0.484 to 0.290, and a Fréchet Radiomics Distance proxy from 330
to ≈ 1, while an independent in-domain grader indicates reduced artifact
severity across all seven classes; all 114 challenge validation outputs
pass an exact metadata-preservation check. Code is available at
https://github.com/Krying/LISA_2026-task1-a.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_013.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=2fYQQKaguW
BibTex
@InProceedings{WonJon_A2.5D_MICCAISAT2026,
author = { Won, Jongjun AND Park, Changseon},
title = { { A 2.5D Ordinal Quality Assessment and Metadata-Preserving Residual Restoration Framework for 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}
}