HemaHier: Chain-Conditioned Ordinal Hierarchies for Lineage-Aware Bone-Marrow Cytology

1 Introduction

• Flat classification treats same-lineage maturation slips and cross-lineage confusions equally, while top-1 accuracy and macro-F1 do not reveal their biological severity or downstream interpretive stakes.Small recognition or lineage-assignment errors can distort differential counts and interpretation, particularly for rare but clinically relevant identities.Source: §1 · p. 2
• Prior approaches structure lineage or maturation, but none couples a lineage-consistent fine posterior with continuous chain-specific maturity while separating healthy chains from atypical and off-chain identities.A single global maturity scalar may not fit distinct lineage trajectories or cells without a well-defined maturity position.Source: §1 · p. 2
• HemaHier introduces a shared ontology and chain-conditioned maturity score, and evaluates recognition alongside error severity and maturation quality.The method targets frozen or lightly adapted cytology foundation models.Source: §1 · pp. 2–3

2 Method

• The ontology assigns each fine cell type to one lineage and encodes healthy cell types in expert-curated maturation chains; dysplastic and off-chain types remain classes without maturity targets.This avoids forcing weakly ordered or abnormal cells onto a healthy developmental axis.Source: §2 · p. 3, Figure 1 · p. 3
• A maturity-aware tree distance scores same-chain errors by normalized maturity gap, same-lineage off-chain errors by an intermediate cost, and cross-lineage errors by the maximum.The distance measures error severity and is not used as a training loss.Source: §2 · pp. 3–4
• The backbone’s late semantic and intermediate morphology-preserving features are factorized into lineage context and detail representations, with the latter capturing texture, granularity, and maturity cues.The fine representation concatenates both factors.Source: §2 · pp. 3–4
• A shared maturity head reads a score under a learned chain query, while the coupled posterior combines lineage probability with fine-class probability renormalized within that lineage.This construction keeps fine predictions consistent with predicted lineage.Source: §2 · p. 4
• Training combines posterior likelihood, auxiliary fine cross-entropy, lineage classification, on-chain stage regression, and within-chain margin ranking, with structural weights ramped in sequence.Fine discrimination stabilizes before lineage, maturity, and ranking supervision are added.Source: §2 · p. 4
• At inference, hierarchy-consistent posterior decoding keeps fine class and lineage in agreement; optional maturity-based recalibration favors classes matching predicted maturation position.Models are selected on validation macro-F1 under the hierarchy-consistent posterior.Source: §2 · p. 5

3.1 Setup

• The evaluation harmonizes UKA2, UKA1, and MLLv1 into a shared ontology and uses five-fold stratified cross-validation with shared splits and training settings across methods.The shared frozen backbone isolates prediction-head effects; the study also reports a rank-16 LoRA-adapted setting and standard recognition, severity, and maturity metrics.Source: §3.1 · pp. 5–6, §3.2 · p. 6

3.2 Results

• Under frozen features, HemaHier leads fine and rare-class F1 on UKA2 and MLLv1 (tied on UKA1 macro-F1); with LoRA, it leads accuracy and fine and rare-class F1 across all three datasets.The MLLv1 frozen-backbone gains over the strongest baseline are 8.9 points in fine macro-F1 (63.3 →72.2) and over twenty points in rare-class F1 (38.6 →59.1).Source: Table 1 · p. 6, §3.2 · p. 7
• With a frozen backbone, HemaHier makes fewer and milder errors, with the lowest cross-lineage error and tree distance across datasets.Fig. 2 decomposes fine-classification error from adjacent-stage maturation slips to cross-lineage confusions; the reduction in severe errors is especially visible on MLLv1 and UKA1.Source: §3.2 · p. 7, Figure 2 · p. 7
• On healthy on-chain cells, predicted maturity rises with true stage (UKA1 neutrophils: Spearman ρ = 0.92), with correlations of ρ = 0.87–0.95 across chains.This continuous within-chain ordering is unavailable from a flat classifier.Source: §3.2 · p. 8, Figure 3 · p. 8

3.3 Component ablation

• On UKA2, factorization, the lineage-consistent posterior, and chain-conditioned maturity progressively lower cross-lineage error and tree distance, while ranking adds a small recognition gain without further improving those severity metrics.The reported changes are 3.6 →3.0 for cross-lineage error and 5.7 →4.9 for tree distance; hierarchy-consistent Bayesian product decoding further lowers both severity metrics without retraining while preserving recognition.Source: §3.3 · p. 8, Table 2 · p. 9

• Across three datasets, HemaHier achieves competitive recognition, particularly strong rare-class performance on the most imbalanced dataset, and lower cross-lineage error and tree distance at a fixed recognition budget.Source: §4 · p. 9
• HemaHier recovers a calibrated within-lineage maturity ordering that flat classifiers cannot express.Source: §4 · p. 9
• Its hierarchy-consistent decoder can trade toward lineage safety at inference without retraining.Source: §4 · p. 9
• The method relies on expert-curated maturation chains, with pathological trajectories identified as a future extension target.Source: §4 · p. 9