Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

KneeXNeT: An Ensemble-Based Approach for Knee Radiographic Evaluation

Nicharee Srikijkasemwat, Soumya Snigdha Kundu, Fuping Wu, Bartlomiej W. Papiez

TL;DR

This work addresses the need for automated knee OA severity assessment by evaluating ten pretrained deep learning models on the OAI radiographs, addressing class imbalance with weighted sampling, and exploring two ensemble strategies (majority voting and a shallow neural network) to produce KneeXNet, an algorithm achieving up to 0.72 accuracy. The approach demonstrates that while individual models peak around 0.69, ensemble methods—especially when combining models trained with and without weighting—can yield higher performance and improved explainability via Smooth-GradCAM++ that highlights clinically relevant joint-space features. The study also investigates the impact of class imbalance on the Rare Doubtful class and suggests potential improvements, such as class-merging or refined KL-based targets, to further enhance automated OA grading. Overall, KneeXNet has potential to assist clinicians by providing faster, explainable, automated knee OA assessments, particularly in settings with limited access to radiology expertise.

Abstract

Knee osteoarthritis (OA) is the most common joint disorder and a leading cause of disability. Diagnosing OA severity typically requires expert assessment of X-ray images and is commonly based on the Kellgren-Lawrence grading system, a time-intensive process. This study aimed to develop an automated deep learning model to classify knee OA severity, reducing the need for expert evaluation. First, we evaluated ten state-of-the-art deep learning models, achieving a top accuracy of 0.69 with individual models. To address class imbalance, we employed weighted sampling, improving accuracy to 0.70. We further applied Smooth-GradCAM++ to visualize decision-influencing regions, enhancing the explainability of the best-performing model. Finally, we developed ensemble models using majority voting and a shallow neural network. Our ensemble model, KneeXNet, achieved the highest accuracy of 0.72, demonstrating its potential as an automated tool for knee OA assessment.

KneeXNeT: An Ensemble-Based Approach for Knee Radiographic Evaluation

TL;DR

This work addresses the need for automated knee OA severity assessment by evaluating ten pretrained deep learning models on the OAI radiographs, addressing class imbalance with weighted sampling, and exploring two ensemble strategies (majority voting and a shallow neural network) to produce KneeXNet, an algorithm achieving up to 0.72 accuracy. The approach demonstrates that while individual models peak around 0.69, ensemble methods—especially when combining models trained with and without weighting—can yield higher performance and improved explainability via Smooth-GradCAM++ that highlights clinically relevant joint-space features. The study also investigates the impact of class imbalance on the Rare Doubtful class and suggests potential improvements, such as class-merging or refined KL-based targets, to further enhance automated OA grading. Overall, KneeXNet has potential to assist clinicians by providing faster, explainable, automated knee OA assessments, particularly in settings with limited access to radiology expertise.

Abstract

Knee osteoarthritis (OA) is the most common joint disorder and a leading cause of disability. Diagnosing OA severity typically requires expert assessment of X-ray images and is commonly based on the Kellgren-Lawrence grading system, a time-intensive process. This study aimed to develop an automated deep learning model to classify knee OA severity, reducing the need for expert evaluation. First, we evaluated ten state-of-the-art deep learning models, achieving a top accuracy of 0.69 with individual models. To address class imbalance, we employed weighted sampling, improving accuracy to 0.70. We further applied Smooth-GradCAM++ to visualize decision-influencing regions, enhancing the explainability of the best-performing model. Finally, we developed ensemble models using majority voting and a shallow neural network. Our ensemble model, KneeXNet, achieved the highest accuracy of 0.72, demonstrating its potential as an automated tool for knee OA assessment.

Paper Structure

This paper contains 14 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Method
  3. Dataset
  4. Baseline models
  5. Ensemble models
  6. Results
  7. Experiment 1: Comparison of the baseline models
  8. Experiment 2: The baseline models with the weighted sampling strategy
  9. Ensemble model from Experiment 1
  10. Ensemble model from Experiment 2
  11. Model explainability using Smooth-GradCAM++
  12. Discussion and Conclusion
  13. Acknowledgements
  14. Compliance with ethical standard

Figures (5)

  • Figure 1: Distribution of samples across each class in the training, validation, and test sets.
  • Figure 2: Test accuracy for each model in Experiment 1 and Experiment 2.
  • Figure 3: F1 scores of each model in Experiment 1 (the baseline models).
  • Figure 4: F1 scores of each model in Experiment 2 (the weighted sampling strategy).
  • Figure 5: Visualizations of the regions influencing model predictions using Smooth Grad-CAM++. These heatmaps highlight the areas of the knee X-ray that the model focuses on when making predictions.