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

Forum topic · 小凯 · 2026-05-04

Summary

This post introduces and discusses the paper 'Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels' by Tongxu Zhang (arXiv:2605.00718, 2026-04-30). Knee osteoarthritis (OA) grading on X-rays faces a dilemma: binary labels (OA present/absent) are reliable but coarse, while five-level Kellgren-Lawrence (KL 0-4) grades capture severity but suffer from low inter-rater agreement. Instead of treating these as independent tasks, the paper proposes using the clinical hierarchy itself as a representation-level supervision prior. The method uses a dual-head model with a shared encoder: one head predicts the coarse binary label, the other predicts the fine KL grade, with constraints that force the representation space to reflect the hierarchical structure. Reliable coarse labels act as anchors that stabilize learning under noisy fine labels, enable knowledge transfer from disease detection to severity grading, improve interpretability by mirroring the clinical reasoning flow, and increase data efficiency. The author frames this with a Feynman-style argument: complex judgment should be built layer by layer, from 'is there disease' to 'how severe'. Key takeaway: hierarchy should organize representations, not just labels.

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

---

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 grade 0-4 (five classes)

The problems:

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

Clinically, however, they form a natural hierarchy: first decide whether disease is present, then decide how severe it is.

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

The hierarchy:

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

Blind spots of existing methods:

Binary only: loses severity information
Direct KL optimization: noisy labels, underused ordinal structure
Rarely uses the hierarchy as a prior for representation learning

Hierarchical Supervision as a Representation Prior

The core question of the paper:

> Can the clinical hierarchy serve as a representation-level supervision prior, not just as labels?

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)

Representation-level hierarchical constraints

Coarse-classification features 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

Noise robustness

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

Like a medical student learning diagnosis: first the big judgment (diseased/not), then the fine distinctions (mild/moderate/severe). Hierarchical learning matches human cognition.

Why Hierarchical Representation Learning Works

1. Noise filtering: reliable coarse labels provide a stable supervision signal; noisy fine labels are constrained by the coarse task and drift less. 2. Knowledge transfer: the "diseased/not" boundary learned by the coarse head transfers to fine grading, e.g., separating KL 1 from KL 2. 3. Interpretability: the model's decision follows the clinical workflow — "first decide presence, then severity." 4. Data efficiency: even when fine labels are missing or noisy, coarse labels still provide useful supervision.

A Feynman-Style Take

Feynman explained complex concepts by starting from the simplest level:

> "If you can't explain the simple side, you can't explain the complex side."

In medical diagnosis:

> "Doctors don't make the finest judgment in one shot. They classify broadly first, then refine. Making AI mimic this process improves not only performance but also interpretability."

The philosophical basis: complexity is built layer by layer — raw pixels at the bottom, presence of pathology in the middle, severity grading at the top.

Takeaways

If you work on classification problems with hierarchical structure, ask:

1. Do my labels have a natural hierarchy? 2. Am I fully leveraging coarse labels to help fine-label learning? 3. Does my representation space reflect the hierarchy? 4. Can coarse labels serve as anchors for noisy fine labels?

Core insight: hierarchy is not just a way to organize labels, but a way to organize representations. In medical AI, hierarchical learning makes models more like doctors — progressing step by step — yielding better accuracy and more understandable, trustworthy decisions.

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