Researchers Propose CoReFuse-Med to Fix Corruption and Modality Bias in Medical Image Fusion

The paper is titled “When Fusion Fails: Corruption-Aware Rebalanced Fusion for Multi-Modal Medical Image Segmentation.” Its authors are Yuchen Pei, Xiaoyu Hu, Yixiong Zou, Dingwen Hu, Hui Chu, Yutao Ma, Shijun Qiu and Gang Li, from Central China Normal University, Huazhong University of Science and Technology, Guangzhou University of Chinese Medicine and its First Affiliated Hospital, and the University of North Carolina at Chapel Hill. On the EPVS dataset, the report said, high-resolution T1 alone reached a DSC of 0.703, while fusing T1, T2 and FLAIR lowered it to about 0.695.

The authors trace the problem to shallow contamination and deep modality bias. A feature contrast-to-noise ratio, F-CNR, measured separability by dividing the difference between lesion and background means by the background standard deviation. In the reported feature visualization, T1 scored 1.0283, T2 0.4301 and FLAIR 0.2980; T1 was about 3.4 times FLAIR. In the 3D U-Net used on EPVS, T1 contributed 77.4% of gradient contribution and FLAIR 13.1%, yet occluding FLAIR reduced DSC by 70.2%, as much as occluding T1. Skip connections in the U-shaped network also carried resampling artifacts into the decoder, the report said, creating feature contamination and greedy modality selection.

CoReFuse-Med keeps a three-stream, symmetric U-shaped segmentation network but divides fusion into shallow cleaning and deep rebalancing. LACE builds anisotropic encodings for each MRI sequence using local DoubleConv and strip-convolution branches along the D, H and W directions, then combines them with optional channel attention. IMSH uses parameter-free 3×3×3 average pooling to separate a structural base from a high-frequency residual; the base uses competitive Softmax weights to select reliable anatomy, the residual is filtered by a shared Sigmoid gate, and a lightweight DSConv3D integrates the parts. MBCF targets L3-L4 deep skip features: it concatenates global descriptors from T1, T2 and FLAIR, passes them through a shared MLP for channel weights, applies spatial gating, and forms a shared query Q that does not belong to any single modality for symmetric cross-attention over concatenated, downsampled K and V.

On EPVS, the method reported DSC 0.7326 and HD95 7.35, compared with MedNeXt at 0.7085 and 8.17. On BraTS it reported DSC 0.8528 and HD95 4.43, compared with MedNeXt at 0.8314 and 7.59. On WMH it reported DSC 0.80, HD95 6.22 and IAVD 0.171 with 2.665M parameters. In an EPVS ablation, a baseline reached DSC 0.7146, while the combination of LACE, IMSH and MBCF reached 0.7322 as Ours(Base). The report said the three modules complemented one another.

The authors also tested BraTS with 20% and 30% of axial slices retained and retrained each model for each setting; the best DSC and HD95 under the two settings were not obtained by applying one set of weights across degradation levels. Additional Gaussian noise was used as a supplementary check. The paper limits its scope to quality loss caused by resolution differences after spatial alignment, not to misalignment or complete modality missing, and not to all strong-noise scenarios. In severely degraded regions of EPVS, the report said, the model mainly prevents contamination from damaging T1 rather than recovering information already lost in T2 or FLAIR. Future work is directed at dynamic modality missing and more acquisition artifacts.