Abstract
Reconstructing healthy tissue in a voided brain MRI is hardest
in the interior of large voids, where the missing structure is underdetermined
by the surrounding context. To address this, we propose a
retrieval-augmented wavelet di usion model that conditions the reconstruction
on an estimate of the missing tissue constructed from other
subjects. For each case, we retrieve the most similar real brains and average
the nearest donors into a prior, and the prior is concatenated to
the wavelet-domain input and ne-tuned end to end. To improve the
generation of healthy tissue, we further use lesion-free brains from the
Human Connectome Project both as additional training data and for the
retrieval donor pool. On a held-out split of the BraTS 2026 inpainting
data, the merged prior improves over a strong unconditioned di usion
baseline, reaching an SSIM of 0.879; a size-resolved analysis shows that
the residual error is concentrated in the interior of large voids. Our results
indicate that a merged retrieved-donor prior is a promising strategy to
supply the non-redundant structure that context alone cannot recover.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Inpainting_019.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=xMP3iml8xh
BibTex
@InProceedings{PerSan_RetrievalAugmented_MICCAISAT2026,
author = { Persson, Sanna AND Bendazzoli, Simone AND Moreno, Rodrigo},
title = { { Retrieval-Augmented Wavelet Diffusion for Local Synthesis of Healthy Brain Tissue } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}