Research map / Label-efficient and robust learning / Recognition and long-tail

Coherent Hierarchical Multi-Label Learning to Defer for Medical Imaging

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Partial and multi-label learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies learning to defer over hierarchical multi-label taxonomies, showing naive per-node L2D yields deferral incoherence and proposing exact coherent projection (DP decoding) and Taxonomic Belief Propagation with Recursive Policy Optimisation (TBP+RPO) that drive incoherence near zero on CXR benchmarks.

Problem

Learning to defer in hierarchical multi-label settings can produce incoherent ternary (predict/defer) action vectors, such as taxonomic contradictions, delegation violations, and deductive defects, which are invisible to flat L2D and standard HMLC.

Approach

Formalises deferral coherence under a Selective-Exclusion handoff contract, gives an exact tree dynamic-programming projection onto coherent action vectors, and proposes a TBP transition model over actions trained end-to-end via RPO through the deployed recursion.

Key results

• On VinDr-CXR, BR+RPO reaches BalAcc 0.853 and F1-L 0.768 with near-zero incoherence (Any 0.004) while BR-L2D has Any 0.107.
• On CheXpert, BR+RPO improves BalAcc to 0.841 vs 0.811 for BR-L2D with Any incoherence ~0.
• BR+Projection gives exactly zero incoherence on all four datasets while staying utility-competitive with BR-L2D.

Surprising findings

• Continued BR training does not consistently reduce incoherence and can increase it (BR cont. Any 0.138 vs BR-L2D 0.107 on VinDr-CXR).

Limitations

• Only tree taxonomies studied; DAG extension discussed only in appendix.
• Evaluation uses retrospective annotations and controlled simulated experts rather than prospective deployment.
• Residual incoherence remains for the fast marginal decoder of TBP+RPO; exact MAP decoding trades off utility.

Experimental setup

Datasets: VinDr-CXR CheXpert PadChest ADPv2

Benchmarks: VinDr-CXR CheXpert PadChest ADPv2

Metrics: balanced accuracy F1-I F1-L taxonomic contradiction rate deductive defect rate delegation violation rate any incoherence rate

Baselines: BR-L2D BR (continued training) BR+Projection

Builds on: Learning to Defer Binary Relevance hierarchical multi-label classification

Closest papers by meaning

• Adaptive Labeling for Efficient Out-of-distribution Model Evaluation (NeurIPS 2024 · similarity 0.86)
• A Two-Stage Learning-to-Defer Approach for Multi-Task Learning (ICML 2025 · similarity 0.84)
• From Biased Selective Labels to Pseudo-Labels: An Expectation-Maximization Framework for Learning from Biased Decisions (ICML 2024 · similarity 0.84)
• Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.83)
• Bootstrapping Chest CT Image Understanding by Distilling Knowledge from X-ray Expert Models (CVPR 2024 · similarity 0.83)
• M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.83)