Neural network-based arterial diameter estimation from ultrasound data.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Cardiovascular diseases are a leading cause of mortality.
• Early detection of carotid artery abnormalities via ultrasound is crucial for prevention.
• Automated carotid diameter waveform extraction is essential for hemodynamic analysis but challenging due to physiological variability.

Purpose of the Study:

• Develop data-driven machine learning (ML) models for automated carotid diameter extraction from ultrasound data.
• Minimize computational complexity for deployment in embedded systems.
• Track carotid artery diameter without requiring clinician annotation or handcrafted heuristics.

Main Methods:

• A ML pipeline using two convolutional neural network (NN) models and a smoothing filter was developed.
• The first NN detects the region of interest (ROI), and the second NN tracks the arterial diameter.
• Ultrasound signals were acquired at 500Hz, and the ML pipeline was trained using a digital signal processing (DSP)-based approach as reference.

Main Results:

• The ML pipeline achieved near-perfect temporal alignment with the DSP reference waveform (Pearson correlation coefficient r = 0.87).
• The mean absolute deviation of arterial diameter prediction was 0.077 mm (1% error).
• The NN-based approach demonstrated robustness against drift and artifacts.

Conclusions:

• The proposed ML pipeline offers a fully automated, accurate, and computationally efficient method for carotid artery diameter tracking.
• This approach eliminates the need for specialist intervention and manual fine-tuning, unlike current clinical practices and conventional DSP methods.
• The ML pipeline's trainability on small datasets and suitability for A-mode ultrasound frames make it promising for miniaturization and on-line clinical/ambulatory monitoring.