Automated Osteophyte Grading

• Osteophyte grading is a standardized process that categorizes bony outgrowths in OA imaging using scales like OARSI and Kellgren-Lawrence.
• Automated methods leverage CNNs and geometric deep learning to improve grading reliability, reproducibility, and diagnostic performance.
• Advanced preprocessing including ROI extraction, intensity normalization, and data augmentation underpins robust deep learning-based osteophyte quantification.

Osteophyte grading is the process of categorizing the size and morphological extent of osteophytes—marginal bony outgrowths—typically for characterizing osteoarthritis (OA) severity in imaging studies. Osteophyte severity is a core radiographic feature in established OA grading systems, notably the Osteoarthritis Research Society International (OARSI) atlas and, in composite form, the Kellgren-Lawrence (KL) score. Recent advances in automated grading leverage deep learning and geometric approaches applied to 2D radiographs and 3D shape data, aiming to improve reliability, reproducibility, and scalability for both research and clinical settings.

1. Osteophyte Grading Scales and Standards

Osteophyte grading is standardized by the OARSI radiographic atlas. For each knee compartment, osteophyte size is graded as an ordinal variable:

Grade	Description
0	No osteophyte
1	Small
2	Moderate
3	Large

Assessment is typically performed at four anatomic sites: medial femur (MF), lateral femur (LF), medial tibia (MT), and lateral tibia (LT), enabling fine-grained OA phenotyping. Inference-time algorithms predict one label per site, allowing independent site-specific quantification. This approach supports both nuanced disease characterization and structured machine learning outputs (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026).

2. Imaging Modalities and Preprocessing

Osteophyte grading is most commonly performed on anteroposterior (AP) knee radiographs, though advanced methods also employ 3D MR-derived bone segmentations. Digital preprocessing protocols include:

• ROI Extraction: Landmarks are detected (random forest, U-Net++, etc.), regions are cropped (typically 140 mm or 140×140 px depending on input resolution), and the image is aligned horizontally with deskewing to standardize view geometry (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026).
• Intensity Normalization: Histogram clipping and global contrast normalization are standard; adaptive histogram equalization (CLAHE) is selectively used depending on pipeline stage.
• Data Augmentation: Random cropping, geometric transformations (rotations, affine scaling), and intensity jittering are used to expand sample diversity, crucial for robust deep learning models (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026).

3D shape analysis begins with MR segmentation, templates, and point correspondence, followed by embedding subject-specific surface deformations for morphometric analysis (Tycowicz, 2019).

3. Algorithmic Approaches to Automatic Grading

3.1 CNN-Based Image Grading

Modern pipelines utilize deep CNN backbones, commonly SE-ResNet-50 or SE-ResNeXt50-32×4d, pretrained on ImageNet and fine-tuned for osteophyte tasks:

• Multi-task Framework: A shared convolutional core is followed by independent FC “heads,” each outputting class logits for a site- or feature-specific grade (e.g., FO-L, TO-L, FO-M, TO-M) (Tiulpin et al., 2019). Gaussian soft-label smoothing can improve ordinal stability (Irfan et al., 25 Apr 2026).
• Loss & Optimization: Each task’s head is trained with cross-entropy (or an ordinal variant), summed across all heads. Adam optimizer with staged learning rate schedules is typical. Balanced mini-batching by class frequency addresses class imbalance (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026).
• Prediction: The argmax of per-site logits yields discrete OARSI grades.

3.2 Geometric Deep Learning for Shape-Based Grading

Semi-supervised graph convolutional networks (GCNs) operate on intrinsic 3D shape representations:

• Pipeline: Shapes are embedded by differential coordinates (deformation gradients), reduced to low-dimensional geodesic sub-manifolds (PGA), and arranged by intrinsic Riemannian distances into a k-NN graph. Node features propagate via multi-layer Chebyshev GCNs, supporting label smoothing and ordinal awareness (Tycowicz, 2019).
• Labeling: Osteophyte categories are typically collapsed to 3 classes (no, minute, definite).
• Transductive Setup: Both labeled and unlabeled subject graphs participate in training, enabling propagation of label information even with limited annotated data.

4. Evaluation Metrics and Quantitative Benchmarks

Performance is assessed by metrics sensitive to ordinal grading and clinical requirements:

• Cohen’s κ and Quadratic Weighted κ: Agreement with expert annotation is quantified via Cohen's κ (unweighted or quadratic, depending on pipeline and dataset) (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026, Tycowicz, 2019).
• ROC AUC/AP: For binary “osteophyte present” (grade ≥1), ROC-AUC and average precision summarize discriminative ability.
• Balanced Accuracy (BA): Accounts for class imbalance in multi-class settings (Irfan et al., 25 Apr 2026).
• Benchmark Results (Selected):

Site (X-ray CNN)	κ (95% CI)	AUC	AP
FO-L	0.79	~0.98	~0.98
FO-M	0.94	~0.98	~0.98
TO-L	0.84	~0.98	~0.98
TO-M	0.83	~0.98	~0.98

Site (Knee-xRAI)	κ (test)	AUC	Bal. Acc.
Med. femur	0.5828	0.6872	0.4427
Lat. femur	0.1048	0.5623	0.3228
Med. tibia	0.4706	0.7105	0.4539
Lat. tibia	0.3665	0.6773	0.4551
Mean	0.3812	0.6593	0.4186

• Interpretation: κ≥0.80 is “excellent” agreement; highest-performing CNN ensembles meet or exceed typical human inter-rater reproducibility. Lower κ for lateral femur sites reflects imaging ambiguities and data imbalance (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026). Shape-based GCNs yield moderate κ (0.58), outperforming Euclidean baselines (Tycowicz, 2019).

5. Workflow Variations, Ensembling, and Model Limitations

• Ensembling: Top-performing models average softmax probabilities across multiple architectures and training “snapshots” (e.g., 10-member ensemble in (Tiulpin et al., 2019)).
• Hyperparameter Selection: Stratified cross-validation and held-out independent test sets (e.g., MOST, OAI) are required to verify performance.
• Site-Specific Tuning: Additional head-only fine-tuning is required for underperforming sites with low ROI counts, especially lateral femur (Irfan et al., 25 Apr 2026).
• Data Scope: X-ray CNNs leverage >20,000 labeled knees for robust generalization; image preprocessing, augmentation, and class rebalancing are crucial. Shape-based pipelines remain dataset-limited (201–300 subjects); this constrains generalizability and necessitates further validation (Tycowicz, 2019).
• Limitations: Single-view 2D grading suffers from projection ambiguity. Lack of multi-reader agreement reporting impedes direct benchmarking against clinical practice. Current shape-based models are transductive; extending to inductive inference and adding joint-space narrowing descriptors would better enable full OA grading (Tycowicz, 2019).

6. Insights, Clinical Interpretation, and Prospects

State-of-the-art osteophyte grading algorithms demonstrate that:

• OARSI-scale site-specific grading is feasible at high reproducibility with CNNs trained on large, expert-labeled datasets (Tiulpin et al., 2019, Irfan et al., 25 Apr 2026).
• Geometric deep learning on intrinsic 3D representations offers moderate accuracy gains and strong theoretical grounding, with limitations in practical workflow integration (Tycowicz, 2019).
• Explainable AI pipelines that decompose canonical OA features (JSN, osteophytes, sclerosis) and structure outputs support auditable, modular grading—and clarify the individual contribution of each phenotype to composite severity (Irfan et al., 25 Apr 2026).

Future improvements are expected from: increasing per-site annotation volumes (especially underrepresented mild and lateral grades), incorporating multi-view or multimodal imaging, augmenting architectures with attention or contrastive learning modules, and systematically quantifying expert-to-expert variability for more rigorous benchmarking. Automated osteophyte grading constitutes a critical component in the broader translation of AI methods toward reliable, large-scale, and explainable OA image phenotyping.