Motion grading of high-resolution quantitative computed tomography supported by deep convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Subject motion in high-resolution peripheral quantitative computed tomography (HR-pQCT) degrades image quality, affecting bone density and morphology measurements.
• Current operator-based scoring of motion artifacts is time-consuming and subjective, potentially leading to acceptance of unusable scans.
• Automated image analysis using Convolutional Neural Networks (CNNs) offers a faster, operator-independent solution for image classification tasks.

Purpose of the Study:

• To develop and validate a CNN model for predicting motion scores in HR-pQCT images.
• To enable automated quality assessment of HR-pQCT scans, identifying uncertain predictions for manual review.
• To improve the precision and reproducibility of bone analysis from HR-pQCT data.

Main Methods:

• Development of a CNN model for automated motion score prediction from HR-pQCT images.
• Evaluation of the CNN's performance using F1-score, precision, recall, and Cohen's kappa agreement.
• Comparison of CNN performance against human operators on a test dataset.

Main Results:

• The CNN achieved a high F1-score (86.8 ± 2.8%) and substantial agreement (Cohen's kappa = 68.6 ± 6.2%) with ground truth motion scores.
• CNN prediction accuracy, precision, and sensitivity were comparable to those of human operators (p > 0.05).
• The CNN processed scans and calculated motion scores within seconds, significantly reducing quality assessment time.

Conclusions:

• The developed CNN provides an accurate and efficient automated method for assessing motion artifacts in HR-pQCT scans.
• This approach can be integrated into clinical workflows to expedite quality control and enhance the reliability of bone microstructural analysis.
• Automated motion scoring by CNNs ensures operator independence and reduces the risk of accepting poor-quality scans, improving patient imaging outcomes.