Abstract
Quality control for magnetic resonance imaging (MRI) usually focuses on one task: catching bad scans. In low-resource pediatric settings, however, a flagged scan often cannot simply be re-acquired, so being able to improve a marginal scan matters as much as being able to detect what is wrong with it. We address both halves of this problem with two independently developed models, sharing only a common cohort of ultra-low-field (0.064 T) Hyperfine SWOOP pediatric brain MRI, a portable, sedation-free scanner increasingly used where conventional high-field MRI is unavailable. Our first model screens each scan for seven common artifact types (Noise, Zipper, Positioning, Banding, Motion, Contrast, Distortion) on a three-level severity scale (none, mild, severe). Because severity is ordered rather than merely categorical, we use a CORN (Conditional Ordinal Regression for Neural networks) ordinal loss applied to a multi-head 2.5D ResNet18 classifier, trained across five patient-level splits with a post-hoc calibration step that corrects the decision thresholds. Our second model attempts to fix, rather than only flag, artifacts – noise, motion blur, zipper-like ghosting, banding, and contrast loss – using a residual U-Net trained to restore artificially degraded versions of real scans, since no genuinely paired clean/degraded acquisitions exist for the same patient. On the official challenge test set the detection model reaches an accuracy of 0.837 and a weighted F1 of 0.808 (micro F1 0.837, macro F1 0.576; patient-level cross-validation gave an accuracy of 0.833 and a weighted F1 of 0.832, with a composite score of 0.836 after calibration. Weighted metrics transfer closely between the two, while macro metrics do not, showing where the ceiling on rare, clinically important cases currently sits. The enhancement model reaches PSNR 33.9 dB, SSIM 0.895 and LPIPS 0.046 on an internal validation cohort under fixed severity-1 noise and motion degradation, and a Fréchet Inception Distance of 124.4 and Fréchet Radiomics Distance of 7.5 under official evaluation on a separate external cohort of real scans. The two models are trained and evaluated separately; we outline, but do not evaluate, how they could be composed into an acquire–assess–enhance–reassess quality-control loop for point-of-care low-field pediatric imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_017.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=FHjWaXHuhW
BibTex
@InProceedings{SidRam_Deep_MICCAISAT2026,
author = { Siddiqui, Rameez Ur Rehman AND Jabbar, Samiya},
title = { { Deep Learning Models for Automated Quality Assessment and Artifact Reduction in 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}
}