Answers are grounded in this paper and cite their sources.
Medical image grading is often ambiguous, particularly between adjacent severity levels. Partial-label learning (PLL) can represent this ambiguity during training, but existing methods generally ignore grade order and still return a single prediction. We propose Ordinal Conformal Partial-Label Learning (OCP-PLL), which converts PLL classifier outputs into contiguous ordinal prediction intervals using an ordinal-aware nonconformity score. With the classifier fixed independently of calibration data, the split-conformal construction provides finite-sample marginal coverage under exchangeability. We also introduce Ord, a fixed Gaussian kernel encoding grade proximity, and UCOR, a learnable unimodal kernel that adapts the ordinal prior during training. Experiments compare 21 methods across nine medical grading datasets using three random seeds. The benchmark includes candidate labels derived from genuine expert disagreement and an eight-grade paediatric bone-age task. Across three PLL families, UCOR improves all five aggregated point-prediction metrics over its corresponding baselines, with an overall macro-accuracy gain of 1.27 percentage points ( p = 0.0117 ). OCP-PLL combines improved grading performance with contiguous severity intervals whose aggregate empirical coverage is close to the nominal 90% target. Code is available at https://github.com/PyJulie/ocp-pll .
No takes yet. Share an insight, caveat, or question.
Ju et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: