Abstract
Ultra-low- eld (ULF) magnetic resonance imaging at 0.064T brings portable, a
ordable pediatric neuroimaging to settings where high-
eld scanners are unavailable, at the cost of low signal-to-noise ratio, weak tissue contrast, and heightened motion sensitivity. We present a single, uni ed deep-learning solution to all three tasks of the LISA 2026 challenge, which spans the ULF acquisition and analysis pipeline. For Task 1a (image-quality assessment) we cast ordinal, multi-label grading of seven artifact families on a three-point severity scale as a slice-based classification cation problem, using an ImageNet-22k ConvNeXt-Tiny backbone with a compact ordinal head trained under a focal + Earth-Mover’s- Distance objective with imbalance-aware sampling. For Task 1b (image enhancement) we train, without paired clean/corrupted data, a residual U-Net restorer adversarially against a PatchGAN discriminator, driving learning with a physically motivated k-space degradation simulator that injects noise, phase-encode motion, ghosting, and spikes on the y. For Task 2 (segmentation of eleven deep brain structures) we build a 3D SegResNet pipeline with deep supervision, anatomy-aware left/right augmentation, test-time augmentation, ve-fold ensembling, and largestconnected- component clean-up, together with a numerical-stability x (b oat16 automatic mixed precision) that resolves an otherwise irrecoverable training failure. On the ocial validation phase our submissions attain a micro-averaged accuracy of 0.835 for Task 1a, a Frechet Inception Distance of 126.0 with a Frechet Radiomic Distance of 9.27 for Task 1b, and a mean Dice of 0:82 0:06 with a mean 95% Hausdor distance of 1.82mm for Task 2. We report all design choices, hyperparameters, and validation-phase results in full.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_023.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=zpYraxHxE6
BibTex
@InProceedings{BaiUjj_Overcoming_MICCAISAT2026,
author = { Baid, Ujjwal AND Baheti, Bhakti AND Madabhushi, Anant},
title = { { Overcoming Low Signal and Scarce Data: Quality Assessment, Enhancement, and Segmentation 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}
}