Development and Validation of Machine Learning Models to Classify Artery Stenosis for Automated Generating Ultrasound Report

Abstract

Duplex ultrasonography (DUS) is a safe, non-invasive, and affordable primary screening tool to identify the vascular risk factors of stroke. The overall process of DUS examination involves a series of complex processes, such as identifying blood vessels, capturing the images of blood vessels, measuring the velocity of blood flow, and then physicians, according to the above information, determining the severity of artery stenosis for generating final ultrasound reports. Generation of transcranial doppler (TCD) and extracranial carotid doppler (ECCD) ultrasound reports involves a lot of manual review processes, which is time-consuming and makes it easy to make errors. Accurate classification of the severity of artery stenosis can provide an early opportunity for decision-making regarding the treatment of artery stenosis. Therefore, machine learning models were developed and validated for classifying artery stenosis severity based on hemodynamic features. This study collected data from all available cases and controlled at one academic teaching hospital in Taiwan between 1 June 2020, and 30 June 2020, from a university teaching hospital and reviewed all patients’ medical records. Supervised machine learning models were developed to classify the severity of artery stenosis. The receiver operating characteristic curve, accuracy, sensitivity, specificity, and positive and negative predictive value were used for model performance evaluation. The performance of the random forest model was better compared to the logistic regression model. For ECCD reports, the accuracy of the random forest model to predict stenosis in various sites was between 0.85 and 1. For TCD reports, the overall accuracy of the random forest model to predict stenosis in various sites was between 0.67 and 0.86. The findings of our study suggest that a machine learning-based model accurately classifies artery stenosis, which indicates that the model has enormous potential to facilitate screening for artery stenosis.

Artificial Intelligence Innovations in Cerebrovascular Neurosurgery: A Systematic Review of Cutting-edge Applications

“…For Transcranial Doppler (TCD), the accuracy of random forest models to predict stenosis ranged from 0.67 to 0.86. The study thus indicated that machine learning-based models accurately classify artery stenosis [57].…”

“…The high accuracy metrics support AI as an adjunct tool in stroke diagnostics [16], [24]. Machine learning models have been developed and shown promising outcomes in classifying stenosis, diagnosing moyamoya disease, brain arteriovenous malformations, cerebrovascular thrombosis [53], [55], [57], [58]. Various neural networks can aid in accurate and quick diagnosis & segmentation of cerebral aneurysms, a task which can be labour intensive if done manually [21], [31].…”