Source-linked AI summary
CATCH: Counterfactual Anatomical Tissue Inpainting with Conditional Haar Diffusion
Simon Winther Albertsen, Hjalte Bjoernstrup, Said Djafar Said, Mostafa Mehdipour Ghazi
TL;DR
Healthy-tissue inpainting aims to replace tumor-distorted regions in brain MRI while preserving observed anatomy. CATCH uses conditional 3D Haar-wavelet diffusion and compares healthy-mask policies; the selected weighted-mixture pipeline outperformed compute-matched tumor-component augmentation on the internal model-selection set and was evaluated on the official validation set.
Problem
Tumors distort brain anatomy and can compromise downstream pipelines, motivating plausible tumor-free inpainting that preserves observed voxels.
Source: §1 · p. 1Method
Results
Weighted mixture outperformed compute-matched tumor-component augmentation on the 75-case internal model-selection set; its selected pipeline achieved SSIM 0.772, PSNR 20.89 dB, and MSE 0.0098 on the official 219-case validation set.
Source: §6 · p. 9, §7 · p. 10Takeaways & Limitations
Within CATCH, the evidence supports selecting the complete weighted-mixture policy rather than attributing the result to its proportions or any single component.
Source: §6 · p. 9, §7 · p. 10Only the weighted-mixture pipeline received official evaluation, and no comparable external baseline was available, limiting claims about its ordering against other pipelines or prior methods.
Source: §6.1 · p. 10
Abstract
from arXiv · showhide
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$\pm$SD) of $0.80\pm0.13$, $19.18\pm1.80$dB, and $0.010\pm0.005$. As the sole officially evaluated pipeline, weighted mixture yielded $0.772\pm0.119$, $20.89\pm3.27$dB, and $0.0098\pm0.0054$ 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.
1 Introduction
The paper frames healthy-tissue inpainting as a way to replace tumor-distorted regions while preserving observed voxels for downstream processing. CATCH uses conditional 3D wavelet diffusion to study how artificial healthy-tissue holes are sampled.
1 Introduction
- Tumors can distort anatomy and compromise registration, parcellation, segmentation, and atlas pipelines; healthy-tissue inpainting replaces pathological regions with plausible tumor-free anatomy.BraTS local synthesis completes a masked region in a voided T1-native volume while preserving observed voxels.Source: §1 · p. 1
- CATCH adapts conditional wavelet diffusion to study artificial-hole sampling, addressing the spatial and morphological limits of supplied masks.The study compares fixed masks, tumor-component augmentation, and a weighted mixture of tumor-derived, irregular-blob, and ellipsoidal masks.Source: §1 · p. 2
2 Related Work
CATCH is situated among diffusion approaches for conditional generation, volumetric synthesis, and brain inpainting, while focusing its comparisons on mask representation and sampling rather than numerical comparisons with prior scores.
2 Related Work
- CATCH retains full ancestral sampling and studies mask representation, tumor-excluded supervision, online healthy-hole sampling, and trajectory aggregation; differing data and evaluation preclude direct numerical comparison with prior scores.The related-work discussion contrasts RePaint's known-pixel restoration with Palette's direct observed-image conditioning, which CATCH also uses before hard compositing.Source: §2 · p. 2
3 Method
CATCH trains a conditional 3D Haar-wavelet diffusion model to inpaint healthy tissue while excluding pathological tumor from supervision. It compares fixed and online healthy-mask policies, then samples and composites reconstructions while preserving observed voxels.
3 Method
- CATCH decomposes the complete inpainting region into disjoint healthy and pathological components, using the complete image as a valid target only in healthy tissue.At inference, the model receives only the voided image and complete inpainting mask.Source: §3 · p. 2
3.1 Conditional Wavelet Diffusion
- A fixed invertible one-level 3D Haar transform produces eight subbands; the denoiser predicts clean coefficients from noisy targets, voided-image coefficients, and the signed mask.The process uses a 1000-step forward schedule and ancestral reverse sampling.Source: §3.1 · pp. 2–3
3.2 Healthy-Tissue Objective
- Tumor-intersecting wavelet cells are excluded from reconstruction error, while a Gaussian-smoothed healthy-hole coverage map weights a regional loss added to the fixed objective.Fixed-grid normalization keeps excluded cells in the denominator, implicitly downweighting per-voxel training signal for cases with larger tumors.Source: §3.2 · p. 4
3.3 Healthy-Mask Training Variants
- Training compares supplied fixed masks, tumor-component-only random augmentation, and a weighted mixture of tumor components, irregular blobs, and ellipsoids.The weighted mixture uses 80/15/5% target proportions; the audit reports observed proportions of 82.2/14.0/3.8% and no fallback.Source: §3.3 · p. 4, Figure 2 · p. 5
3.4 Sampling and Compositing
- Inference uses EMA weights and hard-composites generated tissue inside the mask with the observed image outside it.Random and weighted-mixture pipelines average five case-stable trajectories from one checkpoint, whereas the fixed-mask pipeline uses N = 1.Source: Figure 1 · p. 3, §3.4 · p. 5
4 Experiments
The experiments use development cases to select checkpoints and inference policies, then compare frozen training-policy pipelines on a separate internal set. Evaluation scores the supplied healthy submask and uses paired statistical tests for comparisons.
4.1 Data and Split Protocol
- The 100-case hold-out is split into 25 development and 75 model-selection cases, with 64 model-selection cases patient-disjoint from optimization in a post-hoc analysis.The split is case-keyed: 3/25 development and 11/75 model-selection cases have sibling timepoints in the optimization pool.Source: §4.1 · pp. 5–6
4.2 Checkpoint, Inference, and Training-Policy Selection
- Five prespecified development cases select checkpoints, all 25 development cases select trajectory policies, and the frozen pipelines are compared on 75 separate cases.Checkpoint diagnostics rank checkpoints by aggregate SSIM, PSNR, and MSE; training-policy selection uses the lowest mean joint rank across those metrics.Source: §4.2 · p. 6
4.3 Evaluation and Statistical Analysis
- Evaluation reports SSIM, PSNR, and MSE on the supplied healthy submask, with paired Wilcoxon tests and Holm correction for pairwise contrasts.Means and sample standard deviations describe case heterogeneity, not training-seed variability.Source: §4.3 · p. 6
5 Results
The weighted-mixture pipeline was selected for organizer evaluation after leading the internal frozen-pipeline comparison; development results also determined its checkpoint and trajectory policy. The official validation score reports that pipeline alone.
5.1 Development Selection
- Fixed-mask step 180k was retained after tying step 170k for lowest mean rank; random step 290k and weighted-mixture step 200k led all aggregate metrics within their histories.Fixed step 180k led in PSNR and MSE.Source: §5.1 · p. 7
- On development cases, mean N = 5 had the lowest joint rank for both augmented models, while fixed mask retained N = 1.Fixed-mask mean N = 3 improved rank over N = 1 by only 0.133 at triple trajectory cost, and mean N = 5 was worse.Source: §5.1 · p. 7
5.2 Internal 75-Case Training-Policy Selection
- Against fixed mask, weighted mixture improved SSIM by 0.041, PSNR by 1.34 dB, and [0, 1]-scale MSE by 0.0038, and was selected for organizer evaluation.The primary comparison was not compute-matched because the selected fixed pipeline used N = 1.Source: §5.2 · p. 7
- Against compute-matched random augmentation, weighted mixture improved SSIM by 0.019, PSNR by 0.95 dB, and MSE by 0.0026; all three contrasts remained significant after Holm correction.Fig. 3 presents paired effects across the 75 internal cases.Source: §5.2 · p. 8, Figure 3 · p. 8
- With all pipelines evaluated at mean N = 5, weighted mixture beat fixed mask in 66/75 cases on each metric, so unequal trajectory count did not explain its advantage.The weighted-versus-fixed contrasts had Holm-adjusted p < 2.5×10−11.Source: §5.2 · p. 8
- Weighted mixture won 54/75 cases for SSIM and 57/75 for PSNR and MSE against random augmentation, with gains across all metrics in 51 cases.It lost all three metrics in 15 cases and had mixed effects in 9.Source: §5.2 · p. 8
- In the exploratory hole-size analysis, SSIM and MSE gains were smaller for small scored holes; medium–large differences were not interpreted.The stratification used tertiles of healthy scoring-mask volume, excluding the complete inpainting mask.Source: §5.2 · p. 8
- On the 219-case official BraTS 2026 validation set, weighted-mixture mean-N = 5 achieved SSIM 0.772 ± 0.119, PSNR 20.89 ± 3.27 dB, and MSE 0.0098 ± 0.0054.This organizer result is not a three-arm comparison.Source: §5.2 · p. 8
6 Discussion
The weighted-mixture pipeline was selected for organizer evaluation after leading the internal model-selection comparison, but those cases do not provide independent post-selection confirmation. The discussion also notes inference-compute, split, and clinical-validity limits on interpreting the results.
6 Discussion
- Weighted mixture was strongest on the 75-case internal model-selection set and selected for organizer evaluation, but those estimates are not independent post-selection confirmation.The three frozen pipelines were compared on the same cases used to select the organizer-evaluation pipeline.Source: §6 · p. 9
- At mean N = 5, weighted-vs-fixed advantages persisted, removing unequal inference compute as an explanation while N = 1 remained the primary fixed-mask policy.The frozen fixed-mask pipeline used N = 1 because mean N = 3 gave only a small development-rank gain at triple cost and mean N = 5 was worse.Source: §6 · p. 9
- Mean N = 5 took 29 minutes per case versus 5.8 minutes at N = 1; faster wavelet inpainting should be evaluated under the same cohort and protocol.Times were measured on NVIDIA A100 40 GB GPUs; mean N = 5 required five 1000-step trajectories.Source: §6 · pp. 9–10
- Evaluation covers artificially removed healthy voxels, so metrics and references do not establish patient-specific counterfactual correctness in real tumor regions; clinical validity remains untested.The observed T1n remains pathological in the tumor region, and plausible reconstructions may still hallucinate incorrect anatomy.Source: §6 · p. 10
- The case-keyed split leaves optimization-pool sibling timepoints in 3/25 development and 11/75 model-selection cases, though the patient-disjoint subset retains the pipeline ordering.For 64 patient-disjoint model-selection cases, weighted/random/fixed SSIM was 0.810/0.791/0.770, PSNR was 19.27/18.33/17.93 dB, and MSE was 0.0092/0.0117/0.0128.Source: §6 · p. 10
6.1 Limitations
- One seed per arm, five-case single-trajectory checkpoint selection, transferred inference policies, and no comparable external baseline limit broader claims; only weighted mixture received official evaluation.The 219-case official evaluation cannot confirm the three-arm ordering; fixed-mask mean-(N=5) and void-size analyses were post hoc.Source: §6.1 · p. 10
7 Conclusion
CATCH combines conditional 3D Haar-wavelet diffusion with tumor-excluded supervision, hole-focused loss weighting, and hard compositing for healthy brain-tissue inpainting. Within CATCH, weighted-mixture augmentation led the internal comparison, and its selected mean-N = 5 pipeline was evaluated on the official validation set.
7 Conclusion
- CATCH combines full-volume wavelet-domain diffusion, tumor-excluded supervision, hole-focused loss weighting, and hard compositing of the observed image.Source: §7 · p. 10
- Weighted-mixture augmentation outperformed compute-matched tumor-component augmentation on the 75-case internal model-selection set, with the same ordering in the patient-disjoint analysis.Source: §7 · p. 10
- 0.772 SSIM, 20.89 dB PSNR, and 0.0098 MSE were achieved by the selected weighted-mixture mean-N = 5 pipeline on the official 219-case validation set.Source: §7 · p. 10