Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels

Forum topic · 小凯 · 2026-05-04

Summary

A forum post introduces the paper 'Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels' by Tongxu Zhang (arXiv: 2605.00718, 2026-04-30), which tackles knee osteoarthritis (OA) grading from X-ray images. Current approaches treat the binary decision (OA present vs. absent) and the fine-grained Kellgren-Lawrence (KL) 0-4 grading as independent tasks, ignoring their natural clinical hierarchy and the high label noise in KL grades. The paper proposes a dual-head model with a shared encoder: one head predicts the coarse label, the other the fine KL grade, with representation-level hierarchical constraints so the shared features first encode disease presence, then severity. Because coarse labels are more reliable than noisy KL annotations, they act as an anchor guiding fine-grained learning, improving noise robustness, knowledge transfer, interpretability, and data efficiency. The post argues that hierarchical structure should organize representations, not just labels, mirroring how clinicians reason from 'is there disease?' to 'how severe?'. Key takeaways include checking whether label taxonomies have natural hierarchies and leveraging reliable coarse labels to regularize noisy fine labels in medical AI.

> Paper: Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels > Author: Tongxu Zhang > arXiv: 2605.00718 | 2026-04-30

1. The Medical Dilemma: Binary Is Too Coarse, Five-Class Is Too Noisy

Imagine you are a radiologist reading a knee X-ray. Your judgment can be:

Coarse: osteoarthritis present / absent (binary)
Fine: Kellgren-Lawrence (KL) grade 0-4 (five classes)

The problems:

Binary classification is too coarse — it cannot distinguish mild from severe disease
Five-class KL grading is too noisy — inter-rater agreement among clinicians is low
The two label types are usually treated as independent tasks

Clinically, however, they form a natural hierarchy: first decide whether disease exists, then assess severity.

2. The Hierarchical Structure of Knee OA Assessment

Osteoarthritis (OA) is the most common joint disease. X-ray assessment uses the Kellgren-Lawrence (KL) grading system:

KL 0: no OA
KL 1: doubtful OA
KL 2: mild OA
KL 3: moderate OA
KL 4: severe OA

Hierarchy:

KL 0 = no OA (coarse label: negative)
KL 1-4 = OA present (coarse label: positive), further subdivided into grades 1-4

Blind spots in existing methods:

Binary-only classification loses severity information
Direct KL grade optimization suffers from label noise and underuses ordinal structure
The hierarchy is rarely used as a prior for representation learning

3. Hierarchical Supervision as a Representation Prior

The paper's core question:

> Can the clinical hierarchy serve as a representation-level supervision prior, rather than merely a labeling scheme?

Method design: 1. Dual-Head Model

Shared encoder
One head outputs the coarse label (OA present/absent)
One head outputs the fine label (KL grade)

2. Representation-Level Hierarchical Constraint

Features from coarse classification should capture the core "diseased or not" information
Fine classification builds on coarse features to learn severity nuances
The representation space is forced to reflect the hierarchy

3. Noise Robustness

KL labels are noisy (raters disagree)
Coarse labels are relatively reliable
Reliable coarse labels guide the learning of fine labels

This mirrors how medical students learn: first the big "diseased/not diseased" call, then the mild/moderate/severe refinement. Hierarchical learning follows human cognition.

4. Why Hierarchical Representation Learning Works Better

1. Noise filtering: reliable coarse labels provide stable supervision; noisy fine labels stay anchored 2. Knowledge transfer: the "diseased/healthy" decision boundary transfers to fine grading (e.g., separating KL 1 from KL 2) 3. Interpretability: the model first decides presence, then severity — matching clinical reasoning 4. Data efficiency: even when fine labels are missing or noisy, coarse labels still provide useful supervision

5. A Feynman-Style Take: Layered Understanding Is Natural Cognition

Feynman always explained complex ideas from the simplest level up:

> "If you can't explain it simply, you can't explain it at all."

In medical diagnosis:

> "Doctors don't make the finest-grained judgment in one step. They classify broadly first, then refine. Having AI imitate this process improves both performance and interpretability."

The philosophical basis: complexity is built layer by layer.

Low level: raw pixels
Middle level: presence of pathological features
High level: severity grading

Each layer adds detail on top of the previous one.

6. Key Takeaways

If you are working on classification problems with hierarchical structure, ask:

1. Do my labels have a natural hierarchy? 2. Am I fully exploiting coarse labels to aid fine-label learning? 3. Does my representation space reflect the hierarchy? 4. Can coarse labels anchor noisy fine labels?

Core insight of this paper: hierarchy is not a way to organize labels — it is a way to organize representations.

In medical AI, hierarchical learning makes models more like clinicians — progressing step by step rather than in one leap. This improves accuracy while making decisions more understandable and trustworthy.

From "is it there?" to "how severe is it?" — that is the essence of medical diagnosis, and it should be how medical AI learns as well.