BraTS local synthesis replaces masked regions in T1-weighted brain MRI with plausible tumor-free tissue while preserving observed anatomy. We present CATCH, conditional 3D diffusion in an invertible Haar-wavelet domain. Its denoiser receives noisy target coefficients, voided-image coefficients, and a signed mask; tumor-excluded wavelet reconstruction and a hole-focused loss guide training, and hard compositing preserves observed voxels. We compare fixed masks, tumor-component augmentation, and a weighted mixture of tumor-derived, irregular-blob, and ellipsoidal masks. Of 25 development cases, five prespecified cases select each arm's checkpoint and all 25 of their trajectory aggregations; a separate 75-case internal set compares the frozen pipelines and selects a weighted mixture for organizer evaluation. Five-trajectory averaging yielded internal SSIM/PSNR/MSE (mean±SD) of 0.80±0.13, 19.18±1.80dB, and 0.010±0.005. As the sole officially evaluated pipeline, weighted mixture yielded 0.772±0.119, 20.89±3.27dB, and 0.0098±0.0054 on the 219-case BraTS 2026 validation set. Against compute-matched random augmentation internally, it improved SSIM by 0.019 (95% bootstrap CI: 0.013-0.025), PSNR by 0.95dB, and MSE by 0.003; all three paired comparisons remained significant after Holm correction. Results favor the complete weighted-mixture policy within CATCH; absent official fixed- and random-pipeline scores and a directly comparable external baseline limit broader conclusions.
Figures & tables
Figure 1: CATCH training and inference. A one-level 3D Haar transform produces eight subbands. Noisy target coefficients, voided-image coefficients, and the signed mask form the 17-channel WavUNet input. Predictions are inverse-transformed and hard-composited with the observed image. Random and weighted-mixture pipelines average five case-stable noise trajectories from one exponential-moving-average (EMA) checkpoint; fixed mask uses N=1 .
Figure 2: Training-mask audit. (a) Supplied and generated masks; cyan and orange denote supervised voxels and excluded tumor. (b) Epoch-0 hole-volume distributions (10th–90th percentile, interquartile range, and median; n=100/500/500 ). (c) Target and observed weighted-mixture proportions: tumor/blob/ellipsoid 80/15/5% and 82.2/14.0/3.8% (411/70/19); no fallback occurred.
Configuration
Fixed mask ↓
Random aug. ↓
Weighted mixture ↓
N=1
2.920
4.707
4.400
Mean N=3
2.787
2.133
2.693
Median N=3
2.973
3.533
3.773
Mean N=5
3.240
1.600
1.587
Median N=5
3.080
3.027
2.547
Table 1: Development-set joint rank at the terminal EMA states. Bold and underlining mark the lowest and second-lowest values within each arm. Fixed-mask N=1 was retained for the compute–performance trade-off.
Pipeline
SSIM ↑
PSNR (dB) ↑
MSE ↓ ( ×10−2 )
Rank ↓
Fixed ( N=1 )
0.756 ± 0.162
17.84 ± 1.75
1.365 ± 0.807
2.551
Fixed (mean N=5 )
0.756 ± 0.162
17.85 ± 1.75
1.364 ± 0.807
2.551
Random (mean N=5 )
0.778 ± 0.140
18.24 ± 2.19
1.252 ± 0.718
2.076
Weighted (mean N=5 )
0.797 ± 0.131
19.18 ± 1.80
0.987 ± 0.536
1.373
Table 2: Internal 75-case training-policy selection results (mean ± SD). Lower joint rank is better. Bold and underlining mark the best and second-best results per column. The N=1 pipeline is not compute-matched to the N=5 pipelines.
Figure 3: Compute-matched paired effects of weighted mixture versus random augmentation across the 75 internal model-selection cases. Positive values favor weighted mixture. Points represent cases; diamonds show mean paired effects with 100,000-resample case-bootstrap 95% CIs. W/L denotes wins/losses.
Figure 4: Weighted-mixture mean- N=5 reconstructions at the empirical 10th, 50th, and 90th SSIM percentiles. Cyan denotes the complete inpainting region and dashed yellow the scored healthy region. Error is shown only within the scored region on a shared clipped scale.
Image inpainting reconstructs missing or corrupted image regions using surrounding context, learned visual priors, and optional structural or reference guidance.