Automatic quality assessment of knee radiographs using knowledge graphs and convolutional neural networks

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• X-ray radiography is crucial for diagnostics, but manual quality control (QC) is inefficient and inconsistent.
• Current QC methods struggle with large datasets, necessitating automated solutions.
• Effective QC is vital for maintaining diagnostic accuracy in medical imaging.

Purpose of the Study:

• To develop a multi-criteria artificial intelligence (AI) system for automated knee radiograph quality assessment.
• To enhance the efficiency and consistency of image quality evaluation in radiography.
• To establish AI-driven quality assessment criteria for meeting current QC needs.

Main Methods:

• Developed a knee radiograph QC knowledge graph with 16 acquisition technique and five clarity labels.
• Utilized a ResNet model for simultaneous classification (defect detection) and regression (clarity scoring).
• Trained and validated the AI model on 4324 knee radiographs, with testing on 865 images.

Main Results:

• The AI system (QC4) achieved 98.42% precision for acquisition technique features, outperforming manual assessments (QC1-QC3).
• AI clarity evaluation (QC4) showed a Mean Absolute Error (MAE) of 0.303 ± 0.018, comparable to technologists (QC3 MAE: 0.237 ± 0.016).
• The AI system demonstrated high accuracy in identifying image quality defects.

Conclusions:

• The AI system effectively classifies acquisition techniques for knee radiographs.
• While image clarity evaluation requires further improvement, AI performance is close to that of human experts.
• AI-driven QC using knowledge graphs and convolutional neural networks holds significant potential for clinical application.