SlimFormer++, described in a new arXiv preprint, asks where computation is actually needed in volumetric segmentation networks. The authors analyze block-level redundancy and stage-wise token-mixer contribution, looking for parts of the network that add little representational value.
To guide this analysis, they use Centered Kernel Alignment, which measures similarity between input and output representations. High similarity suggests that a block or stage may be redundant, allowing the network to skip or reduce work there.
The result is a token-mixing design that concentrates computation on the most informative regions of a 3D medical image. The work is positioned as a step toward more efficient 3D medical image segmentation, though the abstract does not yet report specific benchmark numbers.