Computational methods to automate the initial interpretation of lower extremity arterial Doppler and duplex carotid ultrasound studies

Abstract

BackgroundLower extremity arterial Doppler (LEAD) and duplex carotid ultrasound studies are used for the initial evaluation of peripheral arterial disease and carotid stenosis. However, intra- and inter-laboratory variability exists between interpreters, and other interpreter responsibilities can delay the timeliness of the report. To address these deficits, we examined whether machine learning algorithms could be used to classify these Doppler ultrasound studies. MethodsWe developed a hierarchical deep learning model to classify aortoiliac, femoropopliteal, and trifurcation disease in LEAD ultrasound studies and a random forest machine learning algorithm to classify the amount of carotid stenosis from duplex carotid ultrasound studies using experienced physician interpretation in an active, credentialed vascular laboratory as the reference standard. Waveforms, pressures, flow velocities, and the presence of plaque were input into a hierarchal neural network. Artificial intelligence was developed to automate the interpretation of these LEAD and carotid duplex ultrasound studies. Statistical analysis was performed using the confusion matrix. ResultsWe extracted 5761 LEAD ultrasound studies from 2015 to 2017 and 18,650 duplex carotid ultrasound studies from 2016 to 2018 from the Indiana University Health system. The results showed the ability of artificial intelligence algorithms and method, with 97.0% accuracy for predicting normal cases, 88.2% accuracy for aortoiliac disease, 90.1% accuracy for femoropopliteal disease, and 90.5% accuracy for trifurcation disease. For internal carotid artery stenosis, the accuracy was 99.2% for predicting 0% to 49% stenosis, 100% for predicting 50% to 69% stenosis, 100% for predicting >70% stenosis, and 100% for predicting occlusion. For common carotid artery stenosis, the accuracy was 99.9% for predicting 0% to 49% stenosis, 100% for predicting 50% to 99% stenosis, and 100% for predicting occlusion. ConclusionsThe machine learning models using LEAD data, with the collected blood pressure and waveform data, and duplex carotid ultrasound data with the flow velocities and the presence of plaque, showed that novel machine learning models are reliable in differentiating normal from diseased arterial systems and accurate in classifying the extent of vascular disease.