Learning Coarse-to-Fine Osteoarthritis Representations under Noisy Hierarchical Labels
| Authors | Tongxu Zhang |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2605.00718 |
| Download | |
| Categories | cs.CV |
Abstract
Knee osteoarthritis (OA) assessment involves a natural but often underused label hierarchy: a coarse binary OA decision and a fine-grained Kellgren--Lawrence (KL) severity grade. Existing deep learning studies commonly treat these targets as separate classification problems, either reducing OA assessment to disease presence or directly optimizing noisy ordinal KL labels. In this work, we ask whether this clinical hierarchy can serve as a representation-level supervisory prior. Rather than introducing a complex architecture, we use a deliberately simple dual-head model with a shared encoder and two task-specific heads as a probe of hierarchical supervision. We compare single-OA, single-KL, and dual-head training across multiple 3D backbones under the same test protocol. Beyond standard classification metrics, we perform paired statistical comparisons, analyze latent severity-axis geometry, and examine saliency overlap with cartilage regions. The results show that dual-head supervision produces backbone-dependent gains, with clear improvements in KL-related metrics for selected backbones. More importantly, the gains are accompanied by a more ordered coarse-to-fine latent organization and, for responsive backbones, stronger anatomical alignment of saliency with cartilage. These findings suggest that even simple hierarchical dual-head supervision can reshape disease representations under noisy coarse/fine labels, providing a useful inductive bias for OA diagnosis and severity grading.
Engineering Breakdown
Plain English
This paper investigates whether the natural clinical hierarchy in knee osteoarthritis (OA) diagnosis—where doctors first determine if OA is present, then grade its severity using the Kellgren-Lawrence (KL) scale—can improve deep learning models when incorporated as a supervisory signal during training. Rather than treating OA presence detection and KL severity grading as independent classification tasks, the authors use a simple dual-head neural network with a shared encoder that jointly optimizes both objectives. They benchmark this hierarchical approach against single-task baselines across multiple 3D CNN backbones on the same test protocol, effectively using the clinical decision hierarchy as a representation-level prior to guide what features the shared encoder learns.
Core Technical Contribution
The core insight is that task hierarchy from clinical practice can act as an implicit regularization signal during training, without requiring architectural complexity. Instead of inventing a new model, the authors deliberately keep the architecture simple—a shared encoder with two task-specific classification heads—to isolate the effect of hierarchical supervision from architectural novelty. This is a probe-style contribution: they show that jointly learning a coarse-grained task (binary OA presence) and fine-grained task (ordinal KL grades) on the same representation can outperform training these as separate models. The novelty lies in demonstrating that this natural task structure from clinical practice, when encoded as dual supervision, acts as a regularizer that improves both tasks compared to single-task learning.
How It Works
The system starts with a 3D backbone CNN encoder that processes volumetric knee imaging data (likely MRI or CT scans) to produce a shared feature representation. This shared representation is then fed into two separate classification heads: one performs binary classification (OA present or absent), while the other performs multi-class ordinal classification over KL grades (typically 0-4 or 0-5). During training, both heads are optimized simultaneously, but the losses are weighted to balance the coarse and fine-grained objectives. The key mechanism is that gradients from both classification tasks flow back through the shared encoder, forcing it to learn representations that satisfy both the coarse OA decision boundary and the fine-grained severity ordering simultaneously. This differs from training separate single-task models where each encoder specializes only to its own objective, potentially missing features important for the complementary task.
Production Impact
In a clinical deployment pipeline, this hierarchical approach reduces the need to maintain and version separate models for OA screening versus severity assessment—a single model with two outputs handles both decisions from one shared encoder, lowering operational complexity and inference latency. For radiologists or automated triage systems, this architecture provides both immediate binary triage (patient has OA or doesn't) and detailed severity information (exact KL grade) in a single forward pass, improving workflow efficiency compared to running two separate inference calls. The dual-head design also improves generalization: because the shared encoder must satisfy both objectives, it's less likely to overfit to label noise in either task individually, which is critical since KL grading is known to be subjective and sometimes noisy in clinical datasets. The trade-off is minimal—slightly increased model capacity and training time due to optimizing two losses, but gains are real in terms of inference efficiency, model maintenance burden, and robustness to noisy labels.
Limitations and When Not to Use This
The paper assumes that the binary OA and ordinal KL tasks truly share useful representational structure, which may not hold equally well across all anatomical regions, disease stages, or imaging modalities—for rare or atypical presentations, the shared encoder might compromise specialization. The approach requires access to both labels (binary OA presence and KL grade) for all training samples, which can be expensive in clinical settings where grading burden limits labeled data; if only one label is available for some samples, the dual-head framework cannot be fully leveraged. The paper evaluates only on presumably single-center or single-population datasets (not stated explicitly), so it's unclear whether the learned hierarchy generalizes across hospitals with different imaging protocols, expert raters, or patient demographics. Additionally, the work doesn't explore how much of the benefit comes from explicit hierarchy supervision versus implicit multi-task regularization—ablations comparing hierarchical weighting schemes versus equal weighting are not described in the abstract.
Research Context
This work sits at the intersection of multi-task learning and medical image analysis, building on decades of research showing that related tasks can improve each other when trained jointly. It directly addresses a real problem in orthopedic AI: most prior deep learning papers either reduce the problem to binary disease detection (discarding severity information) or directly predict ordinal KL grades (ignoring the natural hierarchical structure and treating noisy ordinal labels as equally important). The paper is part of a broader trend in medical AI toward incorporating domain knowledge and clinical hierarchy into model design, following successful applications of hierarchical learning in fine-grained visual classification. This work opens up the direction of leveraging task hierarchies as implicit inductive biases in medical imaging, potentially applicable to other hierarchical diagnostic tasks (e.g., cancer screening then grading, or disease presence then stage assessment).
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::