Abstract
Brain-MRI analysis pipelines often assume healthy anatomy
and can degrade in the presence of pathology. The BraTS 2026 Inpainting
challenge asks models to fill masked regions, while evaluating reconstruction
inside a withheld healthy region. Most prior entries use encoderdecoder
U-Nets that gather context through downsampling, potentially
sacrificing fine spatial detail. We propose RESIN, a three-dimensional,
norm-free adaptation of SRResNet that processes the volume entirely at
the input resolution. We evaluate it using patient-grouped splits, paired
case-level comparisons and the official scorer. Despite using 5.43M parameters
versus the U-Net’s 15.07M, RESIN outperforms the tuned UNet
on SSIM, PSNR and RMSE on both held-out folds and remains better
in every void-size quartile. Because its receptive field spans the scored
regions, these results support full-resolution processing rather than locality
as the source of its advantage. Across paired ablations, annealing the
learning rate to zero is the strongest training intervention, exceeding the
gains from added capacity, test-time augmentation and alternative loss
terms without increasing inference cost. An ensemble of three RESIN
models attains SSIM 0.8436, PSNR 24.657 and MSE 0.00541 on the official
validation set, surpassing the published 2023 and 2024 challenge
winners on all three metrics, though not the 2025 winner.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Inpainting_020.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=Me8VQDN2pN
BibTex
@InProceedings{CapQua_RESIN_MICCAISAT2026,
author = { Cap, Quan Huu},
title = { { RESIN: Residual Synthesis for Healthy-Tissue Inpainting in Brain MRI } },
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}
}