Use of a Pre-Trained Neural Network for Automatic Classification of Arterial Doppler Flow Waveforms: A Proof of Concept

Abstract

Background: Arterial Doppler flow waveform analysis is a tool recommended for the management of lower extremity peripheral arterial disease (PAD). To standardize the waveform analysis, classifications have been proposed. Neural networks have shown a great ability to categorize data. The aim of the present study was to use an existing neural network to evaluate the potential for categorization of arterial Doppler flow waveforms according to a commonly used classification. Methods: The Pareto efficient ResNet-101 (ResNet-101) neural network was chosen to categorize 424 images of arterial Doppler flow waveforms according to the Simplified Saint-Bonnet classification. As a reference, the inter-operator variability between two trained vascular medicine physicians was also assessed. Accuracy was expressed in percentage, and agreement was assessed using Cohen’s Kappa coefficient. Results: After retraining, ResNet-101 was able to categorize waveforms with 83.7 ± 4.6% accuracy resulting in a kappa coefficient of 0.79 (0.75–0.83) (CI 95%), compared with a kappa coefficient of 0.83 (0.79–0.87) (CI 95%) between the two physicians. Conclusion: This study suggests that the use of transfer learning on a pre-trained neural network is feasible for the automatic classification of images of arterial Doppler flow waveforms.