Coarse-to-Fine AI Diagnosis: Teaching Models to Grade Osteoarthritis Severity Under Noisy Hierarchical Labels

Forum topic · 小凯 · 2026-05-04

Summary

Medical AI systems excel at binary diagnoses (disease present or absent) but struggle with severity grading, a critical limitation since medicine is fundamentally about degrees. This post discusses a 2026 arXiv paper (2605.00718) by Tongxu Zhang that addresses knee osteoarthritis grading using the Kellgren-Lawrence (KL) scale, a five-level system known for significant inter-rater disagreement among radiologists. The key insight is that coarse labels (disease vs. normal) are more reliable than fine-grained KL grades, yet conventional approaches treat these label levels as independent problems—either binary classification or noisy five-class grading. The proposed coarse-to-fine learning framework first learns the low-noise disease/normal distinction, then refines predictions into KL grades within the diseased subset, exploiting the ordinal structure of severity levels while explicitly modeling label noise. The post draws a pedagogical analogy (learning to distinguish cats from dogs before distinguishing breeds) and invokes Feynman's view that acknowledging uncertainty is part of knowledge. It closes with practical design questions for medical AI engineers: does your label structure have hierarchy, what is the noise level, and are you modeling it explicitly?

> Paper: Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels > Author: Tongxu Zhang > arXiv: 2605.00718 | 2026-05-01

The "Black-and-White" AI Diagnosis Problem

When a hospital AI examines a knee X-ray, it may confidently report "osteoarthritis"—but falter when asked how severe it is. This reflects a common weakness in current medical AI: they handle yes/no judgments well but are poor at degree-of-severity judgments.

Medicine, however, is fundamentally about degrees. The same "osteoarthritis" label can mean:

Mild: slight joint space narrowing, barely affecting daily life
Moderate: obvious pain requiring medication
Severe: joint deformity, potentially requiring replacement surgery

For patients, "is it present" and "how bad is it" are entirely different questions.

The Kellgren-Lawrence Grading System: Orthopedics' Five-Level Scale

Orthopedic doctors quantify osteoarthritis severity using the Kellgren-Lawrence (KL) grading:

KL 0: Normal
KL 1: Doubtful, possible minute osteophytes
KL 2: Mild, definite osteophytes but normal joint space
KL 3: Moderate, joint space narrowing
KL 4: Severe, marked narrowing with sclerosis

This 60-year-old system is one of orthopedics' most widely used standards. The problem: KL grading is extremely noisy. Different doctors examining the same X-ray often assign different grades, with only moderate inter-rater agreement reported in studies.

This means the "ground truth" AI learns from is itself noisy.

The Coarse-vs-Fine Label Dilemma

The research highlights an overlooked issue: osteoarthritis has two label levels:

Coarse: Disease (OA) / Normal
Fine: KL 0 / 1 / 2 / 3 / 4

Existing deep learning work typically treats these as independent problems—either binary classification or direct five-class grading. The costs:

Binary-only: all severity information is lost
Direct five-class: heavily disrupted by KL label noise

The two label hierarchies are related, complementary, and should be learned jointly.

Coarse-to-Fine: Using Hierarchy to Fight Noise

The paper proposes a coarse-to-fine learning framework:

1. Learn coarse first: distinguish "diseased" vs. "normal"—an easier judgment with less noise 2. Then learn fine: within the diseased subset, further separate KL 2/3/4 3. Hierarchical linkage: exploit the ordinal relation that KL 3 is necessarily more severe than KL 2 to constrain predictions 4. Noise modeling: explicitly model label noise rather than pretending labels are perfect

It's like teaching a student to distinguish cats from dogs first, then distinguishing golden retrievers from huskies within dogs.

The Feynman Perspective: Uncertainty Is Part of Science

Feynman emphasized in his quantum mechanics lectures:

> "We need not be ashamed of admitting uncertainty. Uncertainty is part of knowledge."

Medical AI is no different. When expert agreement on KL grading is only moderate, pretending labels are 100% correct and forcing AI to fit noisy labels is self-deception. The more honest approach:

Explicitly model label uncertainty
Use prior knowledge embedded in the label hierarchy
Let reliable coarse information guide difficult fine-grained decisions

This isn't lowering standards—it's handling real-world messiness intelligently.

Takeaways for Medical AI Designers

If you're building a medical classification system, ask:

1. Does my label structure have a natural hierarchy? 2. What is the noise level of the fine-grained labels? 3. Are coarse labels more reliable and usable to guide fine-grained learning? 4. Am I explicitly modeling label noise?

Medicine isn't black and white. A good medical AI system must judge not only "whether" but also honestly express "how severe"—and how confident it is in that judgment. When AI learns to understand disease from coarse to fine, from certain to uncertain, it truly begins to think like a doctor.