"Radiobiometrics": Deep-learning radiograph biometrics for patient identification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biometrics

Background:

• Accurate patient identification is crucial for radiology follow-up examinations.
• Manual verification is challenging due to patient condition changes and imaging variations.
• Misidentification errors can lead to significant patient safety risks.

Purpose of the Study:

• To develop an automated system for patient identification from radiographs using deep learning.
• To enhance the quality assurance of radiology reporting.
• To explore the potential of radiographs as a biometric modality.

Main Methods:

• Utilized convolutional neural networks (CNNs) for automatic patient identification.
• Employed deep metric learning to train models for matching radiographs.
• Trained models for various body parts (chest, knees, pelvis, hands) and multi-view chest images.

Main Results:

• Achieved over 0.98 true positive rate (TPR) at 0.001 false positive rate (FPR) for most models.
• Attained over 0.96 rank-1 TPR on internal test datasets.
• The multi-view chest radiograph model demonstrated high accuracy in matching frontal and lateral views.

Conclusions:

• The developed CNN system effectively identifies patients from radiographs.
• Radiographs show potential as a reliable biometric modality for subject identification.
• This technology offers significant quality assurance benefits for healthcare institutions.