XGA-Osteo: Towards XAI-Enabled Knee Osteoarthritis Diagnosis with Adversarial Learning

Hieu Phan Trung, Loc Le Tan, Mao Nguyen, Phat Nguyen, Sang Nguyen, Minh-Triet Tran, Tho QuanIJCAI 2024

Introduction

Knee Osteoarthritis (KOA) is a leading cause of global disability, yet its clinical diagnosis remains largely manual and subjective. While deep learning models have shown promise in automating severity grading, they typically suffer from two critical limitations: the "black box" problem, where models provide scores without anatomical justification, and a dependency on massive, expertly labeled datasets for specific pathological features like bone spurs.

XGA-Osteo addresses these gaps by employing unsupervised anomaly detection. Instead of learning to identify disease directly, the system learns the distribution of "healthy" knee anatomy and identifies pathological deviations as anomalies. This approach provides both high-resolution explainability and improved diagnostic performance.

Methodology

The core of the system is the Osteo-GAN framework, which operates in three distinct stages:

  1. ROI-Masked Reconstruction: The system identifies the knee joint's Region of Interest (ROI) using segmentation and blurring to mask the central joint area. The Osteo-GAN, trained exclusively on healthy knee images, is then tasked with "restoring" this masked area as a healthy representation.
  2. Anomaly Map Generation: By calculating the pixel-wise difference between the patient’s original X-ray and the GAN’s healthy reconstruction, the system generates a high-resolution heatmap. Since the GAN only reconstructs healthy features, pathological markers like subchondral sclerosis or joint space narrowing appear as significant differences.
  3. Dual-Input Classification: Unlike standard models, XGA-Osteo uses a two-stream architecture. It processes both the original X-ray and the generated anomaly map through a shared CNN backbone. These features are concatenated to produce the final Kellgren-Lawrence (KL) severity grade.
Figure 1: The XGA-Osteo diagnostic pipeline, illustrating the transition from input image to anomaly map and final diagnosis.

Experimental Setup and Datasets

The researchers utilized data from the Osteoarthritis Initiative (OAI), a multi-center longitudinal study. A specific subset of healthy images was used to train the Osteo-GAN, while the broader dataset served to evaluate classification performance across the full spectrum of disease severity.

Results and Discussion

XGA-Osteo was benchmarked against several state-of-the-art CNN backbones, including ResNet and DenseNet.

  • Improved Sensitivity: While raw accuracy was competitive with baselines (~76-79%), the proposed method significantly boosted Recall (the ability to detect disease). In one test case, recall jumped from 74.43% to 81.33%, suggesting that the anomaly map acts as a critical "hint" for the classification network.
  • Superior Explainability: Traditional explainable AI (XAI) methods like Grad-CAM often produce coarse, blurry blobs. In contrast, XGA-Osteo provides precise anatomical maps that align with clinical markers used by radiologists.
Figure 4: Qualitative comparison between XGA-Osteo anomaly maps and Grad-CAM heatmaps, showing more precise localization of clinical features.

Conclusion and Future Work

XGA-Osteo demonstrates that treating disease as a deviation from a reconstructed healthy norm can bypass the need for expensive manual labeling while improving transparency. By providing clinicians with high-resolution visual justifications, the system moves closer to a collaborative "human-in-the-loop" diagnostic process. Future research will focus on clinical validation with practicing radiologists and refining the GAN architecture to eliminate potential generative artifacts.

Summary updated

BibTeX
@inproceedings{trung2024xga,
  author = {Hieu Phan Trung and Loc Le Tan and Mao Nguyen and Phat Nguyen and Sang Nguyen and Minh-Triet Tran and Tho Quan},
  title = {{XGA-Osteo: Towards XAI-Enabled Knee Osteoarthritis Diagnosis with Adversarial Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure