Luma proposes an exact sequence parallelism method for packed multimodal masking. At each layer, the method chooses among ring attention, sparse all-to-all, or Ulysses, and its numerical results are consistent with those on a single GPU. On up to 128 H100 GPUs, the method outperforms the FlashMask + Ulysses hybrid scheme in 47 of 48 cases. There are currently no follow-up developments.
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 propose a multimodal MRI segmentation network, CoReFuse-Med. The method uses LACE+IMSH to remove resampling artifacts at shallow layers and MBCF to calibrate modality contributions at deep layers, addressing the problem of low-quality modalities contaminating the fusion. On the EPVS dataset, it achieves a DSC of 0.7326 and an HD95 of 7.35, outperforming MedNeXt's 0.7085 and 8.17; on BraTS, its DSC is 0.8528 with only 2.665M parameters. The method targets the resolution degradation after spatial alignment and does not cover scenarios with missing modalities or strong noise.
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 propose a multimodal MRI segmentation network, CoReFuse-Med. The method uses LACE+IMSH to remove resampling artifacts at shallow layers and MBCF to calibrate modality contributions at deep layers, addressing the problem of low-quality modalities contaminating the fusion. On the EPVS dataset,
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 propose a multimodal MRI segmentation network, CoReFuse-Med. The method uses LACE+IMSH to remove resampling artifacts at shallow layers and MBCF to calibrate modality contributions at deep layers, addressing the problem of low-quality modalities contaminating the fusion. On the EPVS dataset, it achieves a DSC of 0.7326 and an HD95 of 7.35, outperforming MedNeXt's 0.7085 and 8.17; on BraTS, its DSC is 0.8528 with only 2.665M parameters. The method targets the resolution degradation after spatial alignment and does not cover scenarios with missing modalities or strong noise.