Automated Ultrasound Doppler Angle Estimation Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Ultrasound

Background:

• Accurate Doppler angle estimation is crucial for reliable blood velocity measurement in ultrasound.
• Incorrect angle estimation is a primary source of error in Doppler-based velocity assessments.
• Current manual methods are prone to human error and variability.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for automated Doppler angle estimation.
• To compare the accuracy of the automated method against human expert measurements.
• To assess the potential clinical impact of automated angle estimation in reducing diagnostic errors.

Main Methods:

• A deep learning model was trained on 2100 human carotid ultrasound images with data augmentation.
• Pre-trained models extracted image features, which were fed into a custom shallow network for angle estimation.
• Automated angle estimates were compared to measurements made by human observers.

Main Results:

• The mean absolute error (MAE) for automated angle estimation ranged from 3.9° to 9.4° across evaluated models.
• The best performing model achieved an MAE below the clinical threshold for acceptable Doppler angle error.
• This accuracy level helps prevent misclassification of normal velocities as stenosis.

Conclusions:

• Deep learning offers a viable technique for automated ultrasound Doppler angle estimation.
• The proposed method demonstrates high accuracy and potential for clinical implementation.
• Integration into ultrasound scanners could enhance diagnostic reliability and reduce errors in blood velocity measurements.