Hierarchical Multi-Label Learning to Defer for Medical Imaging
A new study introduces the first Learning to Defer (L2D) framework for hierarchical multi-label decisions, motivated by medical-imaging workflows where findings are organized by clinical taxonomies. The research, published on arXiv, addresses deferral incoherence issues such as taxonomic contradictions and delegation violations. It formalizes coherent hierarchical deferral under a Selective-Exclusion handoff contract and proposes two remedies: exact coherent projection using a dynamic-programming decoder and Taxonomic Belief Propagation.
Key facts
- First L2D setting with hierarchical multi-label decisions
- Motivated by medical-imaging workflows with clinical taxonomies
- Deferral is a delegation action, not a label assignment
- Identifies deferral incoherence including taxonomic contradictions, delegation violations, and deferrals of implied labels
- Formalizes coherent hierarchical deferral under a Selective-Exclusion handoff contract
- Characterizes the Bayes-optimal coherent deferral rule
- Shows that nodewise Bayes L2D can be action-incoherent
- Proposes exact coherent projection and Taxonomic Belief Propagation as remedies
Entities
Institutions
- arXiv