Abstract
The 2026 BraTS-Inpainting task requires local synthesis of
healthy-appearing tissue in masked T1-weighted brain MRI scans. The
reconstructed image aims to appear anatomically coherent and match the
surrounding intensity distribution, inpainting the masked region without
introducing a visible boundary. In this work, we present a patch-based
3D attention U-Net trained with a mask-normalized objective function
that combines voxel error, multi-scale structural similarity (MS-SSIM),
Haar-wavelet agreement, and inner and outer boundary penalties. The
proposed training framework adds four components to an existing baseline:
(i) a checkpoint-persisted cosine scheduler capable of resuming after stopping,
(ii) exponential moving average (EMA) weights with validation-time
selection between raw and averaged parameters, (iii) interpolation-free
spatial and mild intensity augmentation, and (iv) SSIM-prioritized loss
and checkpoint criteria with a masked-MSE safeguard. During inference,
Gaussian-weighted sliding windows, test-time augmentation, and checkpoint
ensembling improve predictions, while exact compositing leaves
healthy voxels outside the inpainting mask unchanged. On the official
challenge leaderboard, the proposed method achieves a mean SSIM of
0.8138, MSE of 0.0064, and PSNR of 23.41 dB on the online test set.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Inpainting_009.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=KDaKbCLkmt
BibTex
@InProceedings{HosMD_An_MICCAISAT2026,
author = { Hossen, MD Fayaz Bin AND Evans, Michael L. AND Sadique, MD Shibly AND Farzana, Walia AND Rahman, Asfaqur AND Temtam, Ahmed AND Iftekharuddin, Khan M.},
title = { { An SSIM-Prioritized 3D Attention U-Net for Synthetic Brain Inpainting } },
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}
}