Abstract
In many resource-limited imaging centers, MRI studies are archived as unordered JPEG slice collections without spatial metadata, rendering them unusable for modern deep learning pipelines that rely on structured 3D formats (e.g., DICOM or NIfTI). Existing reconstruction methods assume access to k-space data, known acquisition geometry, or intact modality structure. These assumptions break in real-world, low-infrastructure settings. We introduce the first unified framework for metadata-free 3D MRI reconstruction from unstructured, incomplete JPEG archives. Our transformer-GAN pipeline jointly infers anatomical slice ordering, reconstructs missing slices, and generates high-fidelity 3D volumes within a single end-to-end architecture. An attention-based slice tokenizer enables order-invariant processing, a novel anatomy-driven transformer recovers spatial arrangement, and progressive 3D modeling reconstructs the volume, with an adversarial refinement stage that sharpens fine anatomical detail. Experiments on BraTS demonstrate robust slice reordering, accurate missing-slice recovery, and high-quality volumetric reconstruction, approaching the performance of fully ordered and complete inputs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_018.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=nPay2lWdei
BibTex
@InProceedings{IorAon_From_MICCAISAT2026,
author = { Iorumbur, Aondona Moses AND Pennisi, Matteo AND Raymond, Confidence AND Zhang, Dong AND Ibitoye, Ayo Zacchaeus AND Anazodo, Udunna C. AND Spampinato, Concetto},
title = { { From Disorder to Volume: Transformer-Based 3D MRI Reconstruction from Unstructured JPEG Archives } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}