Interpretable and parameter optimized ensemble model for knee osteoarthritis assessment using radiographs.

Area of Science:

• Orthopedics
• Medical Imaging
• Artificial Intelligence

Background:

• Knee osteoarthritis (KOA) significantly impacts patient mobility and quality of life.
• Current KOA assessment using X-rays, focusing on features like osteophytes and joint space narrowing (JSN), is often subjective.
• Objective quantification of KOA severity is crucial for effective patient management.

Purpose of the Study:

• To develop an objective and interpretable AI-driven system for assessing knee osteoarthritis (KOA) severity.
• To improve the accuracy of KOA grading using convolutional neural network (CNN) models.
• To provide visual explanations for AI predictions to enhance clinical trust and utility.

Main Methods:

• Implemented an interpretable ensemble of CNN models for KOA assessment.
• Developed a scale-invariant model for knee joint localization.
• Utilized hyperparameter-optimized CNNs and an ensemble scoring system for Kellgren-Lawrence (KL) grading.
• Validated the models on 37,996 knee joints from the Osteoarthritis Initiative (OAI) dataset.

Main Results:

• Achieved superior performance improvement of 13-27% compared to existing state-of-the-art methods.
• Demonstrated accurate and objective quantification of KOA features from low-resolution X-ray images.
• Provided visual explanations for model predictions, enhancing interpretability.

Conclusions:

• The developed AI system offers a more objective and accurate method for KOA assessment.
• This approach can aid clinicians in prioritizing patient care and treatment.
• The interpretable nature of the model promotes trust and potential adoption in clinical practice.