CoReFuse-Med Cleans and Rebalances MRI Modalities for Segmentation

What happened

CoReFuse-Med, a medical image segmentation method described in a paper by researchers affiliated with Central China Normal University and other institutions, first addresses noisy shallow features and then rebalances modality contributions. On the EPVS dataset, it achieved a DSC of 0.7326 and HD95 of 7.35, compared with MedNeXt’s 0.7085 and 8.17. The study focuses on resolution differences after spatial alignment; it does not address cases where all modalities are missing.

The story so far

Yuchen Pei and co-authors developed CoReFuse-Med, a method for combining MRI modalities in medical image segmentation, and reported an EPVS dataset DSC of 0.7326 and HD95 of 7.35, compared with MedNeXt’s 0.7085 and 8.17. The method first addresses noisy shallow features, then rebalances how much each modality contributes. The study tackles a problem in which adding MRI modalities can worsen segmentation: on the IH EPVS dataset, the authors found that high-resolution T1 alone achieved a DSC of 0.703, while combining T1, T2 and FLAIR lowered it to about 0.695. Their approach focuses on quality loss caused by resolution differences after spatial alignment. In severely degraded EPVS regions, it aims to prevent degraded inputs from contaminating T1 features, rather than recover information already lost from T2 or FLAIR. Its scope does not include unaligned inputs, cases where all modalities are missing, or every severe-noise scenario.

Keep reading