Can convolutional neural networks identify external carotid artery calcifications?

ABSTRACT

OBJECTIVE: We developed and evaluated the accuracy and reliability of a convolutional neural network (CNN) in detecting external carotid artery calcifications (ECACs) in cone-beam computed tomography scans. STUDY DESIGN: Using TensorFlow, we developed a program to identify calcification in 427 cone-beam computed tomography scans evaluated to determine the presence of ECACs. We compared the results to the findings of a human evaluator. Using an 8020 training-to-validation ratio, we calculated the k-fold cross-validation accuracy of the initial dataset and extrapolated the F1 score and Matthews Correlation Coefficient. RESULTS: We calculated a k-fold cross-validation accuracy of 76%, with a recall and precision of 66% and 79%, respectively, and a combined F1 score of 0.72. We extrapolated a Matthews correlation coefficient of 0.53, showing a strong balance between confusion matrix categories. CONCLUSION: Our CNN model can reliably identify ECACs in cone-beam computed tomography scans.