Arterial diameter trend estimation using deep learning on ultrasound spectral doppler

This study presents an approach for estimating gaps in arterial diameter using flow velocity obtained from spectral Doppler data. We utilize short-time Fourier transform in conjunction with deep learning models designed for spectrograms to estimate arterial diameter trends. We train the Convolutional Recurrent Neural Network with Attention (CRNN-A) and Audio Spectrogram Transformer (AST) to provide scaled trend estimates in increments of 1, 2, or 4 seconds (s). We present an algorithm for Arterial ReScaling (AReS) that utilizes scaled trend predictions to fill in missing arterial diameter data at 1, 2 or 4 s increments. The CRNN-A model trained on 1 s lengths performed best at predicting scaled trends (R2 = 0.8083). However, when implementing AReS, the AST model trained on 2 s segments performed best at predicting the shortest gaps (1 s) in arterial diameter data (MAE: 0.0341 mm). Our study showcases that training to predict longer time segments can be beneficial in real-world performance, even if standardized metrics show otherwise.Clinical Relevance - In cases where an arterial ultrasound image becomes temporarily lost or inaccessible, the spectral Doppler signal can be used to estimate changes in the arterial diameter, offering continuous data. This allows for an accurate tracking of both blood flow and volume changes, assisting in the study of vascular reactivity and potentially aiding in clinical diagnoses such as carotid stenosis.