AGE-Net: Spectral--Spatial Fusion and Anatomical Graph Reasoning with Evidential Ordinal Regression for Knee Osteoarthritis Grading

TL;DR

Knee Kellgren--Lawrence grading from radiographs is challenged by subtle cues, long-range anatomical dependencies, and label uncertainty near boundaries. AGE-Net integrates Spectral--Spatial Fusion, Anatomical Graph Reasoning, and Differential Refinement on a ConvNeXt backbone, with an evidential Normal-Inverse-Gamma head and an ordinal ranking constraint to produce calibrated, monotone predictions. The approach achieves state-of-the-art performance on a knee KL dataset (QWK around 0.90 and low MSE) and is supported by ablations showing the value of each component. This framework offers uncertainty-aware, interpretable grading suitable for clinical triage and population-scale screening, with ongoing work on per-grade behavior, uncertainty quality, and robustness.

Abstract

Automated Kellgren--Lawrence (KL) grading from knee radiographs is challenging due to subtle structural changes, long-range anatomical dependencies, and ambiguity near grade boundaries. We propose AGE-Net, a ConvNeXt-based framework that integrates Spectral--Spatial Fusion (SSF), Anatomical Graph Reasoning (AGR), and Differential Refinement (DFR). To capture predictive uncertainty and preserve label ordinality, AGE-Net employs a Normal-Inverse-Gamma (NIG) evidential regression head and a pairwise ordinal ranking constraint. On a knee KL dataset, AGE-Net achieves a quadratic weighted kappa (QWK) of 0.9017 +/- 0.0045 and a mean squared error (MSE) of 0.2349 +/- 0.0028 over three random seeds, outperforming strong CNN baselines and showing consistent gains in ablation studies. We further outline evaluations of uncertainty quality, robustness, and explainability, with additional experimental figures to be included in the full manuscript.

Figures (9)

• Figure 1: AGE-Net overview. A ConvNeXt backbone extracts feature maps, followed by SSF, AGR, and DFR. Global average pooling yields an embedding fed to the COE evidential head to produce NIG parameters. Training optimizes evidential loss plus ordinal ranking.
• Figure 2: Spectral--Spatial Fusion (SSF). SSF performs frequency-domain modulation (rFFT/iFFT with learnable complex weights) and spatial attention recalibration with residual stabilization.
• Figure 3: Anatomical Graph Reasoner (AGR). Features are pooled to a fixed grid, flattened into tokens, and processed via kNN graph + EdgeConv-style aggregation. A sigmoid gate is upsampled and applied to the original feature map with residual connection.
• Figure 4: Differential Refiner (DFR). DFR uses depthwise filtering and absolute feature difference to enhance boundary-sensitive cues.
• Figure 5: COE-Head: evidential NIG regression. The head predicts $(\gamma,\nu,\alpha,\beta)$ and derives uncertainty from evidence parameters.