Clinics often hang T1 with a sparse FLAIR or T2 that only kept a fraction of the axial slices. After you resample everything onto one grid, a multi-modal network can look worse than T1 alone. CoReFuse-Med is a corruption-aware fusion stack that filters resampling junk on the weak series in shallow layers, then rebalances modality weight deeper so the fused overlay does not get pulled around by the noisy path. Yuchen Pei, Xiaoyu Hu, Yixiong Zou, Dingwen Hu, Hui Chu, Yutao Ma, Shijun Qiu, and Gang Li posted arXiv:2609.10261 on 9 September 2026 (ACM MM 2026). Code is linked at github.com/lrever/CoReFuse.
Viewer teams already know the hang: T1 is dense, FLAIR or T2 arrives with roughly one tenth of the Z samples, rigid registration makes the volumes look co-registered, and early fusion still injects aliasing into the skip paths. The paper names that failure mode modality-quality mismatch under spatial alignment, not missing modalities and not large misregistration.
Why ordinary fusion breaks on the aligned but uneven hang
On their in-house enlarged perivascular spaces (EPVS) MRI set, T1 is high resolution while FLAIR and T2 keep about 10% of the axial slices. A standard 3D U-Net multi-modal run lands below the T1-only baseline (about 0.695 DSC versus a higher T1-only score in their Fig. 1 setup). Gradient accounting shows T1 dominating training updates, yet occlusion tests show the weak series still swinging the prediction hard. That split is the optimization-inference inconsistency the authors target: the degraded path is under-trained and over-influential at once.
They attribute the mess to two coupled effects. Resampling puts structured high-frequency noise into the low-quality modality, and U-Net skips carry that noise into the decoder. Under that corruption the optimizer takes a greedy shortcut, leaning on the clean series while still letting the noisy features spoil boundaries. Attention-heavy fusion that assumes clean inputs tends to amplify the problem rather than damp it.
What CoReFuse-Med actually does
The framework splits fusion into two stages instead of one early concatenate-and-attend block.
A Lightweight Axial-Context Encoder (LACE) reads anisotropic volumes with axial-aware convolutions so sparse Z sampling is not treated like isotropic 3D texture. Inter-Modal Spatial-Scale Harmonization (IMSH) runs in shallow layers (roughly L1 to L2). It peels each modality into a smooth anatomical base (3×3×3 average pool) and a detail residual, competes the bases across modalities for a shared layout, then gates the residuals with a shared depthwise-separable path so resampling artifacts get attenuated before they ride the skips. Modality Bias Calibration Fusion (MBCF) sits deeper (L3 to L4). It builds global modality descriptors, predicts channel weights from a joint descriptor, applies spatial gates, and runs a light symmetric cross-attention so no single high-SNR series quietly owns the fusion.
They ship two backbones: Ours (Base) with ordinary DoubleConv blocks, and Ours (Med) with MedNeXt-style inverted bottlenecks. Training uses 96-cubed crops, z-score normalization, and a Dice/Focal mix (0.6 / 0.4) on an RTX 3090. CoReFuse-Med (Med) sits around 2.7M parameters, far smaller than Swin-UNETR in their comparison table.
Where it was measured
Three settings. IH EPVS is the real clinical anisotropy case (T1 dense, FLAIR/T2 sparse). Synthetic BraTS 2020 keeps T1ce full resolution and retains 10% of FLAIR and T2 Z slices (with 20% and 30% retention ablations, plus an auxiliary noise test). WMH Challenge covers white-matter hyperintensity under information loss with the official metrics (DSC, HD95, IAVD, F1).
On EPVS, CoReFuse-Med reaches about 0.73 DSC and 7.35 mm HD95, ahead of early-fusion 3D U-Net, Swin-UNETR, and MedNeXt, while specialized fusion nets (MMFormer, HNoSegXS) drop below 0.55 DSC under the same mismatch. On degraded BraTS, the Med variant averages about 0.85 DSC across WT/TC/ET with HD95 around 4.4 mm, tighter than MedNeXt’s high-variance boundary errors. On WMH it matches the sysu_media leaderboard F1 (0.76) as a single model with HD95 about 6.2 mm. Ablations on EPVS show LACE, IMSH, and MBCF each help, and the full stack is best. Metrics stay as support; the shop claim is fusion that survives a sparse series instead of quietly losing to T1-only.
How this lands in a viewer
If you already hang multi-contrast brain MRI in a DICOM viewer, treat CoReFuse-Med as a fusion policy for uneven Z sampling, not a new worklist. Read spacing and slice counts per series, resample onto the dense T1 (or T1ce) grid, run LACE plus IMSH on shallow features, then MBCF deeper, and hang the EPVS, tumor, or WMH overlay as an editable layer. Fail closed when modalities are missing entirely (the paper does not claim missing-modality completion), when rigid alignment is wrong, when you need full isotropic native resolution on every contrast, or when the sparse series has no recoverable lesion signal and should be dropped from fusion rather than force-fed. The in-house EPVS set stays private; rebuild BraTS and WMH protocols from the public code link.
SpFiLM conditions one parcellation network on a T1 versus T1ce flag. This paper is dense plus sparse multi-contrast in, fused overlay out without letting the weak series spoil the hang.
Rebuild from arXiv:2609.10261. As of 10 September 2026 the abstract and PDF respond (HTTP 200). The preprint is the source of record for the ACM MM 2026 camera-ready. Public code is linked in the paper.
Sources
- Yuchen Pei, Xiaoyu Hu, Yixiong Zou, Dingwen Hu, Hui Chu, Yutao Ma, Shijun Qiu, Gang Li. When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation. arXiv:2609.10261, posted 9 September 2026 (ACM MM 2026). https://arxiv.org/abs/2609.10261 (HTTP 200 on 10 September 2026). PDF: https://arxiv.org/pdf/2609.10261 (HTTP 200 on 10 September 2026). Code: https://github.com/lrever/CoReFuse.