Auto-evaluation of skull radiograph accuracy using unsupervised anomaly detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiography

Background:

• Accurate patient positioning in radiography is crucial for diagnostic image quality.
• Suboptimal radiographs are often rejected, necessitating repeat imaging, which is currently a subjective process.
• Automating the assessment of radiograph quality can enhance efficiency and consistency.

Purpose of the Study:

• To develop and evaluate an automated system for assessing skull radiograph accuracy using unsupervised machine learning.
• To replace qualitative, visual assessment of radiograph suitability with a quantitative, data-driven approach.
• To utilize autoencoder (AE) and variational autoencoder (VAE) models for anomaly detection in radiographic images.

Main Methods:

• Acquired 1,680 skull radiograph images from five phantoms, categorized as normal (appropriate positioning) or abnormal (inappropriate positioning).
• Employed anomaly detection techniques based on AE and VAE models to discriminate between acceptable and unacceptable radiographs.
• Verified the discriminatory capability of the unsupervised learning models on the acquired dataset.

Main Results:

• The autoencoder (AE) model achieved an area under the curve (AUC) of 0.7060 in receiver operating characteristic analysis.
• The variational autoencoder (VAE) model achieved an AUC of 0.6707.
• The proposed unsupervised learning methods demonstrated superior discrimination ability compared to previous studies with 52% accuracy.

Conclusions:

• The developed method exhibits high classification accuracy for determining the need for skull radiograph retakes.
• Automating the assessment of radiograph quality can significantly improve operational efficiency in high-volume X-ray departments.
• This approach offers a quantitative and objective alternative to manual radiograph evaluation.