Source-linked AI summary
LoFi RADIO: A Distilled In-Domain Backbone Applied for Artifact-Severity Grading of Ultra-Low-Field Neonatal Brain MR
Jonathan B. Martin, Yashwant Kurmi, Charlotte R. Sappo
TL;DR
Neonatal ULF MRI is artifact-prone, motivating automated severity grading across heterogeneous scans. The paper evaluates complementary foundation-model routing and in-domain distillation, finding that LoFi-RADIO improves the weighted composite while using one deployable student backbone.
Problem
ULF neonatal MRI artifact grading is challenging because scans have low SNR and other acquisition difficulties and come from multiple low-field scanners.
Method
The study routes complementary foundation-model teachers per artifact and distills selected teachers into a single in-domain ViT-S student.
Results
0.8374 overall composite was achieved after adding Phikon and exact-field OpenNeuro data to DINOv2 and BiomedCLIP distillation.
Takeaways & Limitations
LoFi-RADIO consolidates complementary teacher strengths into one backbone without requiring multiple foundation models or a router at inference.
Takeaways & Limitations
The study uses magnitude-only image-domain data, and evaluation on additional downstream low-field MRI tasks remains to be performed.
Abstract
from arXiv · showhide
Ultra-low-field MRI makes neonatal brain imaging deploy- able in low-resource settings, but its low SNR, lack of shielding, and long scan duration make it especially prone to acquisition artifacts, motivating automated quality control. We address the LISA 2026 Task 1a challenge: multi-label severity grading (0/1/2) of seven common image artifacts on ULF T2 weighted volumes. We identify that a number of backbones may be successfully paired with a classification MLP, but that no single backbone is uniformly best across artifacts. To improve performance, we evaluate routing complementary foundation model teachers through a per-artifact gate, as well as distilling the teachers into a single in-domain ViT-S student (LoFi RADIO) over an unlabeled low-field MRI corpus. Both of these strategies improve the weighted composite. The distilled backbone matches or exceeds the gate and has the added advantage of not requiring deployment of multiple large foundation models at infer- ence.
1 Introduction
ULF neonatal MRI is portable and low-cost but particularly vulnerable to artifacts, creating a challenging multi-label severity-grading quality-control problem across scanners.
- 0.064 T ULF MRI offers a low-cost, portable option for neonatal neuroimaging in resource-limited settings.
- Low SNR, absent shielding, long acquisitions, neonatal motion, and adult RF coils exacerbate image-quality problems.
- Scanner diversity makes the task challenging because image quality and artifact characteristics differ across acquisition sources.
- Complementary foundation models can be distilled into specialized compact students, motivating an in-domain low-field MRI backbone.
2 Methods
The method evaluates frozen backbone probes and multi-backbone alternatives for seven ordinal artifact grades, then distills complementary teachers into a single in-domain student for efficient grading.
- Task and data: Seven artifacts are graded independently per volume on an ordinal scale of 0, 1, or 2.
- Task and data: The dataset contains 531 magnitude-only T2-weighted acquisitions from 243 subjects across three sites and orientations.
- Evaluation: The headline metric averages support-weighted accuracy, F1, F2, precision, and recall composites across the seven artifacts.
- Backbone screening: Frozen candidate backbones are evaluated with identical single-arm probes, subject-grouped folds, tuned heads, and multiple seeds.
- Multi-backbone methods: Multi-backbone strategies compare feature concatenation with a learned per-artifact gate that routes predictions from complementary teachers.
- LoFi-RADIO: LoFi-RADIO distills DINOv2, BiomedCLIP, and Phikon into a ViT-S/14 student trained on unlabeled low-field images.
- LoFi-RADIO: The student reproduces teacher summary and dense spatial representations through adaptor MLPs and standardized teacher targets.
- LoFi-RADIO: SAM is excluded from distillation because its low-rank feature contributes through routing and cannot be reproduced by one fused student.
3 Results
Screening found complementary artifact strengths across frozen backbones, while per-artifact gating and in-domain multi-teacher distillation improved the composite. LoFi RADIO reached a best held-out composite of 0.8322, ranking third among 16 teams.
- Every candidate but ARNIQA fell within approximately 0.013 of DINOv2, while neither degeneracy metric predicted performance.
- The four-backbone gate reached 0.8327, recovering complementary per-artifact strengths that naive feature concatenation did not convert into improvement.Naive concatenation reached 0.8293.
- 0.8374 was the composite after adding Phikon and pooling the exact-field 0.064 T OpenNeuro corpus into distillation.A student distilled from DINOv2 and BiomedCLIP reached 0.8330, while adding Phikon increased the composite to 0.8349 before exact-field data were pooled.
- The exact-field corpus left the overall composite unchanged within noise but redistributed improvement toward weak artifacts and reduced across-seed variance.
- 0.8322 was LoFi RADIO’s best composite on the held-out dataset, ranking third of 16 submitting teams.The two higher reported team scores were 0.8356 and 0.8334.
4 Discussion
The study argues that complementary foundation-model strengths can be consolidated through multi-teacher distillation into one compact in-domain student. It also identifies limitations involving data representation, corpus scope, contrast coverage, and cross-site distribution shift.
- A frozen foundation model was usually a strong feature extractor, but no single backbone was best across all artifacts.
- Artifact-specific routing recovered complementary model strengths better than simple feature concatenation, while multi-teacher distillation consolidated them into one higher-performing student.
- Frozen utility was largely insensitive to pretraining objective, whereas BrainIAC clearly failed because its augmentation-invariance objective discarded appearance variation indicating artifact severity.
- A pathology-pretrained teacher may have contributed most improvement, attributed to fine-grained textural cues shared by some artifacts and histopathology images.
- The magnitude-only, T2-weighted LISA data exclude phase information and do not establish performance across other contrasts or mixed-contrast datasets.
- Pooling the external low-field corpus produced only a marginal within-noise overall change, while a larger and more diverse exact-field corpus may be needed for further improvement.
- The study does not evaluate the distilled backbone on additional downstream low-field MRI tasks such as segmentation or image enhancement.