Deep learning for carotid Doppler spectra classification.

Cardiovascular disease (CVD) remains a global health challenge, with limited specialist access in low- and middle-income countries hindering early detection. Carotid Doppler ultrasound offers promise for screening in non-specialist settings. However, spectral Doppler lacks the anatomical context provided by duplex ultrasound, making it challenging to determine which carotid vessel is being assessed. This study focuses on accurately identifying signals from the common, internal, and external carotid arteries (CCA, ICA, and ECA) based solely on Doppler spectra. This forms a critical step for subsequent disease classification. A clinical study enrolled 398 participants who underwent bilateral carotid Doppler examination (198 healthy controls, 200 with CVD). Several classifiers were evaluated including i) five deep convolutional neural networks (CNN) utilizing transfer learning on spectral images, and ii) conventional machine learning classifiers applied to the maximum frequency envelope and extracted features. The best performing classifier was a CNN (GoogLeNet) which achieved a mean area under the curve (AUC) of 0.929, effectively distinguishing between carotid artery segments. It exhibited f1-scores of 0.830, 0.803 and 0.764 for the ICA, ECA and CCA, respectively. Explainable AI tools (GradCAM and LIME) provide intuitive visual insights into these predictions. This study addresses a previously unsolved problem of automated carotid artery segment identification using spectral Doppler waveforms alone. When incorporated into automated screening tools, this approach provides a low-cost, specialist-independent pathway for earlier CVD detection, particularly suited to resource-constrained environments.