Fig. 5
From: SALT: Introducing a framework for hierarchical segmentations in medical imaging using label trees
Overview of the SALT architecture and process. (A) A DynUNet model outputs feature maps with N channels, where N corresponds to the number of nodes in the hierarchical label tree. The SALT layer then computes conditional probabilities along parent–child paths and replaces the standard softmax for generating structured segmentations. (B) Example visualization of the conditional probability maps used to segment the L4 vertebra. The model does not predict the target class directly but infers it through the hierarchical path: spine → lumbar spine (LS) → L4.
