Abstract

Automated quality control (QC) of ultra-low-field (ULF) neonatal MRI must be learned from only a few hundred annotated volumes with severe class imbalance, and is a prerequisite for trustworthy downstream analysis. We address LISA 2026 Task 1a, grading seven artifact categories—noise, zipper, positioning, banding, motion, contrast, distortion—on an ordinal three-point scale. We show that, in this small-data regime, a simple slice-based 2D transfer-learning pipeline—extracting a fixed number of slices per volume, encoding each with an ImageNet-pretrained backbone, aggregating by mean-and-max pooling, and reading out ordinal heads—substantially outperforms a 3D convolutional network trained from scratch. A Swin-Tiny backbone reaches an internal out-of-fold weighted score of 0.830 versus 0.818 for EfficientNet-B0 and 0.805 for a 3D DenseNet; on the validation leaderboard our 2D systems score about 0.821 versus 0.806 for the 3D system. We add two validated cautions: post-hoc decision-threshold and ensemble-weight optimization is non-transferable on this cohort, rejected by both nested cross-validation and a resampling test sized to the evaluation cohort; and the support-weighted ranking metric discounts the rarest artifact, so a dedicated minority-class model raising banding macro-F1 from 0.32 to 0.74 does not improve—and can slightly reduce—the ranking. We discuss implications for metric-aware model selection.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_005.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=RA2UclSM4d

BibTex

@InProceedings{ManTai_SliceBased_MICCAISAT2026,
        author = { Manabe, Taiki AND Fukuse, Hiroaki},
        title = { { Slice-Based 2D Transfer Learning Outperforms 3D-From-Scratch for Quality Assessment of Ultra-Low-Field Neonatal 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}
}


back to top